DEV Community: Shawn

FutureX · Physical AI Daily — Issue 111 (09/06)

Shawn — Sat, 05 Sep 2026 14:52:15 +0000

Today's Highlights

· Zhejiang University and Alibaba DAMO Academy attach a 40M-parameter corrector to VLA models, lifting disturbance-recovery success rate to 68.3%

· Guangxiang Technology's world model retires after training, hitting 80.3% on LIBERO-PLUS

· AGIBOT (Chinese humanoid startup) and Minth's Serbian plant ships its first batch of robots off the line on September 4

· Nissan's Smyrna body shop deploys AMRs, leaving 64 forklift positions unfilled going forward

· Figure's crowdsourced app collects 16 million videos and has paid creators $15 million

· XDOF emerges from three months of stealth, with Series B valuation talks near $1.2 billion

Research Progress

Fitting a VLA with a 40M-parameter brake for on-the-fly self-correction · vla

A robotic arm is placing a block into a bowl when someone moves the bowl away — the model is still receiving camera input, but its hand keeps executing an action generated hundreds of milliseconds earlier. The open-loop blind spot between action chunks is exactly what the OmniAI team at Zhejiang University's ACES Lab, working with Alibaba's DAMO Academy, targeted: without touching the backbone weights, they attach a roughly 40M-parameter Corrector outside the inference pipeline that compares the expected versus actual change in visual features within latent space; if the deviation persists, it flushes the queue of not-yet-executed actions and turns the deviation into a gradient signal that guides that recovery inference pass. On the real-world AgileX PiPER platform, average success rate across nine tasks rose from 55.6% to 73.3%, and the disturbance-recovery task (where the target is manually moved) rose from 40.0% to 68.3%. The gains go beyond failure rate: under SmolVLA at horizon 10, success rate rose from 61.90% to 73.00%, while the average number of policy calls actually fell from 19.27 to 15.64, with 83.7% of truncations occurring at critical stages such as grasping, alignment, and insertion.

Pan Yi et al. (Zhejiang University · Alibaba DAMO Academy) · arXiv 2607.01804 source · Analysis: Sina Technology source

A world model that only clocks in during training, stepping aside once the robot is at work · world-model

After swapping camera viewpoints, changing lighting and backgrounds, and adding sensor noise, average success rate on LIBERO-PLUS still reaches 80.3%, versus 51.5% for Fast-WAM. Phi-WM 1.0 ActEffect, developed by Guangxiang Technology (Chinese world-model startup) with Professor Li Shengbo's group at Tsinghua University, confines a controlled world model to the training phase: the policy produces three complete candidate actions in one pass — feedforward, coarse proposal, and refined — and the world model predicts the consequences of each within a frozen DINOv3 feature space, comparing each against the real future outcome and requiring that the refined action beat the coarse proposal, which in turn beats the feedforward action; this ranking is then backpropagated into the policy weights. Once training ends, the world model and the future-observation branch are both stripped away, so no future rollout or candidate-action search happens at execution time. On the 29-dimensional-action-space RoboCasa-GR1 benchmark, average success rate reaches 67.5%, 9.2 points above the runner-up ABot-M0; removing the consequence feedback drops LIBERO performance from 98.8% to 97.0%. What's saved is the per-step inference overhead on every unit on a production line.

Guangxiang Technology × Tsinghua University (Li Shengbo group) · Phi-WM 1.0 ActEffect · Analysis: QbitAI source

Formalizing "who goes through the doorway first" gives multi-robot navigation its first unified taxonomy · autonomy

Two delivery robots racing toward the same door at once, several autonomous vehicles arriving simultaneously at an unsignaled intersection, robot fleets meeting head-on in a corridor. This survey, published in Autonomous Robots, names this class of friction a Social Mini-Game, defined as a situation involving two or more agents whose optimal trajectories collide within the same time window such that someone must yield; the authors model it as a partially observable stochastic game and provide geometric criteria for determining when ordinary navigation slides into an SMG. Their diagnosis is that disciplinary fragmentation is what's slowing progress on the "last mile" problem: conflicts that humans resolve gracefully with a small speed adjustment still haven't been replicated in robots.

Rohan Chandra et al. (University of Virginia · UT Austin · UIC · CMU Robotics Institute) · Autonomous Robots · Analysis: Bioengineer.org source

Open Source · Tools · Benchmarks

· ABC Dataset: released jointly by XDOF and UC Berkeley BAIR, which the company claims is the largest high-quality robot training dataset to date, recorded via remote teleoperation and wearable sensors on human operators performing everyday tasks like folding laundry and flattening cardboard boxes source

· VLA-Corrector: Zhejiang University and DAMO Academy released both code and a project page; the 40M-parameter bypass module can be attached directly to π0.5 or SmolVLA without modifying backbone weights source

· Magic-VLA K02: MagicLab (Chinese humanoid startup)'s general-purpose embodied foundation model made its European debut at IFA; the company says it will open-source the base foundation model this October source

Funding & Deals

XDOF | Series B talks | ~$1.2 billion valuation · adjacent ⚠️ Rumored terms

8VC is leading, terms are undecided, and neither the total raise nor whether the valuation includes new money has been confirmed. This real-world teleoperation data company was founded in 2024 by UC Berkeley's Philipp Wu (CEO) and Fred Shentu (CTO), and only closed a $70 million Series A this past June, with participation from Thrive Capital, Andreessen Horowitz, Lux, and Spark Capital. The company hadn't planned to raise again so soon, but annualized revenue approaching $50 million drew investors in. The founders' technical starting point was GELLO, a low-cost teleoperation system they built during their PhDs; investors now compare the company to a robotics-focused Scale AI.Source: TechCrunch source

Quanzhibo | Series B | hundreds of millions of yuan · hardware

Led by Yida Capital and Guotai Haitong Kaiyuan, with participation from Western Securities Equity Investment, just a month after its A+++ round in July — this is the eighth funding round in under two years for this Wuxi-based joint-module maker. The funds go toward ramping up to million-unit annual capacity and developing next-generation high-torque-density joint modules. Shipments topped 100,000 units in 2025, with over 60,000 units shipped in June 2026 alone, and first-half-2026 shipments already exceeding all of last year; automation rate exceeds 85%, with first-pass yield holding above 96%. Lead investor Yida Capital's assessment is that joint modules make up 30%-60% of a whole robot's cost, making them a key link for self-sufficiency in China's robotics supply chain. AGIBOT, LimX Dynamics, and Linker Bot are both customers and shareholders.Source: PEdaily source

Heimo Technology | three rounds in half a year | nearly 100 million yuan total · hardware

Yida Capital took a stake in April and then led a subsequent round; shareholders also include four industrial partners — LimX Dynamics, Dongfang Precision, Zhaofeng Co., and Shenhao Technology. Founder Yang Zihe's starting point for building dexterous hands traces back to the southwest Indian Ocean: during a scientific expedition aboard the research vessel Dayang Yihao, an ROV's two-finger gripper couldn't grasp complex-posed seafloor targets, and after a multi-hour dive the mission returned empty-handed. On September 4 the company released its reconfigurable dexterous hand DaSheng 1 (SG100), with 4 fingers, 11 active degrees of freedom, repeatability precision of up to 0.03mm, and a whole-hand payload of 24kg. Another model, WA100, is described by the company as the world's first waterproof dexterous hand, with the deep-sea engineering version rated for a maximum operating depth of 10,000 meters and a food-grade-material civilian version aimed at kitchen and bathroom scenarios.⚠️ Manufacturer's claimSource: 21st Century Business Herald source

Faor Robotics | A-share IPO guidance filing | sponsor Guotai Haitong · industrial

Registration with the Jiangsu Securities Regulatory Bureau was completed yesterday, with Chairman Yao Ting holding a 24.74% stake. This Suzhou-based company, founded in 2019, focuses on intelligent-sensing collaborative robots and says it is the first cobot maker to self-develop all core components — harmonic reducers, motors, controllers — in-house; it completed its joint-stock restructuring this past March. In 2025 it received 13,000 cobot orders and shipped 11,000 units, including 3,200 units overseas/international; first-half-2026 orders were up 120% year-over-year. Investors across its previous six funding rounds include Shunwei Capital, Hillhouse Ventures, Alibaba, Meituan Long Zhu, Source Code Capital, and China Life Capital.Source: Robotics Outlook source

Commercialization & Deployment

First batch of humanoid robots ships from the Šabac, Serbia plant · humanoid

On September 4, the first batch of robots left the production line. Following last week's start of production, this joint venture between Minth (Chinese auto-parts maker) and AGIBOT — Serbia's first humanoid robot plant — has now completed the step from production start to shipment, with two humanoid models and one quadruped model running on the line. Robots leave the factory as general-purpose hardware; task assignment happens later via AI and specialized software depending on where they're deployed, with industrial applications coming first and home applications later. Factory general manager Djin Mao told RTS that the goal isn't just to build robots but to train local technical personnel; production supervisor Dejan Babic trained for 30 days in Jiaxing, China, participating in trial production and logistics from material intake to finished-goods dispatch. Staffing plans call for over 200 employees. The capacity figures don't quite match between reports: RTS reported a planned annual output of 3,000 units, while the target announced at the start of production was over 5,000 units annually. Minth previously produced automotive parts in Serbia; the local supply chain for sensors, batteries, motors, and chips that the robotics venture is now building up did not previously exist there.Source: Srpske Novine source

Nissan's Smyrna plant deploys AMRs, leaving 64 forklift positions unfilled after attrition · industrial

Each unit carries a payload of about 4,190 pounds at a top speed of 4.5 mph, with a second unit stepping in whenever one falls behind on task progress. The AMRs running in the Smyrna, Tennessee body shop come from OTTO, a Rockwell Automation subsidiary; Rockwell acquired OTTO's parent company Clearpath Robotics for about $600 million in October 2023, and the same model is also used in Ford, GE, Hershey, and Caterpillar plants. Nissan lists this as its largest cost-reduction project this year, one that will ultimately displace 64 forklift and tugger operator positions. Affected employees remain with the company and can apply for other roles, with some potentially retrained to operate industrial robots; but once the transition is complete, these material-handling positions will not be refilled. The software has had glitches too, at one point dispatching two robots to the same location simultaneously, which Nissan had its supplier fix. The rollout has just completed the first of six planned phases.Source: Hoodline, citing Business Insider source

MagicLab makes its European debut at IFA, signs deal with Slovenia's postal service · humanoid

The MagicBot D1 is already handling shop-floor material handling and line loading/unloading at Dreame's smart manufacturing plant — evidence of deployment beyond the trade-show floor. On September 4, MagicLab unveiled three new products in Berlin: the full-size humanoid MagicBot X1, with knee joint peak torque above 350N·m and a limit of 450N·m; the industrial wheeled humanoid D1, with a working-height range of 80 to 230cm and a 7-degree-of-freedom force-controlled arm with 0.5N force-control precision; and the lightweight quadruped MagicDog T1, weighing 19kg with a sustained payload of about 15kg. Commercially, the company has struck a partnership with Slovenia's postal service for flexible sorting at a local distribution center, and the quadruped MagicDog Y1 is being used for inspections at European energy plants and power stations. The company says it holds and is pursuing orders worth over 1.1 billion yuan, expects 500 million yuan in revenue in 2026, aims to raise the overseas/international revenue share from 26% in 2025 to 50%, and expects CE-RED whole-machine safety certification to be completed by Q4.⚠️ Manufacturer's claimSource: iFeng Review source

Galbot (Chinese embodied-AI startup)'s G1 completes an unattended pick-up demo at IFA · embodied

At booth 160 in IFA Next, Hall 25 of Messe Berlin, a G1 unit picked items off a shelf, navigated through crowded space, and manipulated objects with no operator intervention. This 85kg wheeled semi-humanoid stands 173cm tall with a 190cm arm span and a torso reach of 240cm when extended, equipped with dual eye cameras, four torso-mounted RGB cameras, a wrist depth camera, chassis lidar, and eight ultrasonic sensors, plus cloud connectivity for model updates and remote monitoring. European buyers at the booth don't face the hurdle now facing the US market: on July 28, the US FCC added to its covered list any foreign-manufactured mobile robot weighing over 2kg that combines environmental-sensing sensors, wireless connectivity above 200kbps, and autonomous navigation software; new models of such robots are no longer eligible for the equipment certification required for import, marketing, or sale in the US, though previously certified older models are unaffected. The product specs on display at the show match all three criteria in that definition.Source: Tech Times source

Industry Developments

Yu Kai: Horizon Robotics aims to overtake Nvidia in China's high-end autonomous-driving chip market next year · autonomy ⚠️ Stated plan

"As things stand today, we're number two in the industry, right behind Nvidia." At the Yabuli Entrepreneurs Forum on September 5, Horizon Robotics (Chinese autonomous-driving chipmaker) founder and CEO Yu Kai described the company's position in high-compute autonomous-driving chips this way, setting a goal of overtaking Nvidia next year — a plan stated by Yu Kai himself, not yet backed by independent data. According to a poster released by Horizon Robotics, in China's market for autonomous-driving chips in Chinese-brand passenger vehicles, Horizon holds the top spot at 31.94%, with Nvidia close behind at 29.38%; following mass production of the Journey 6 series, its share rose to 22.82%, placing it second among chip suppliers supporting urban NOA (navigate on autopilot). Yu Kai also disclosed that Robot Era (robotics venture he's incubated) was recently valued at over $3 billion in its latest funding round, covering compute spanning robot vacuums, lawn mowers, and humanoid robots — a segment where Nvidia remains the main rival.Source: Yicai source

XPeng unveils X-Mind, embedding a world model inside its large driving model · world-model

Before outputting an action, the system first runs a spatiotemporal simulation "in its head" — this is the visual chain-of-thought X-Mind adds to the in-vehicle agent. The framework uses Recurrent Block Diffusion to progressively denoise across layers within a single forward pass, generating a compact abstract sketch, from which the planner derives the ego vehicle's trajectory based on this predicted physical future. XPeng's reasoning is that mainstream approaches are stuck at a reactive perception-to-action mapping — equivalent to a driver hitting the gas while staring only at the current frame; pure text-based reasoning can't clearly express complex environmental geometry, while directly predicting raw future images introduces a lot of high-frequency redundant texture. At the embodied-AI foundation model deployment workshop at CVPR 2026 in Denver this June, the head of XPeng Group's General Intelligence Center first publicly presented the complete technical roadmap, naming active inference, controllable generation, and long-horizon prediction as the three necessary capabilities for a high-performance world model. In the first half of the year, the team had already published three research reports on world models: X-World, X-Foresight, and X-Cache.Source: XPeng official site source

Figure launches Index, turning ordinary people's household-chore videos into training data · humanoid

Publicly launched on August 25, it has so far received over 16 million videos from 108 countries, with 264,000 downloads and 44,000 weekly active users. The system processes 30 minutes of video per second, equivalent to about 4.9 years of human activity uploaded per day; every 1,000 hours of data covers an average of 373 distinct tasks, 1,146 manipulated objects, and 116 environments. Figure has paid $15 million to contributors it calls Creators, and has pledged to invest over $1 billion to expand data collection and compute, aiming to scale up 100x within a year. Raw videos go through a five-stage cleaning pipeline before feeding Helix, the VLA model powering Figure 03. The project began because third-party data vendors couldn't deliver the diversity and quality Figure wanted. It was developed secretly for four months under the codename Project Go-Big; coming a week after Figure's $3.5 billion compute agreement with Nscale, these two major commitments announced within a single week point to the same bottleneck.Source: Crypto Briefing source

Shenzhen issues 2026–2028 intelligent robotics industry plan · adjacent

Shenzhen's Bureau of Industry and Information Technology and Bureau of Science, Technology and Innovation jointly issued the "Shenzhen Work Plan for Promoting High-Quality Development of the Intelligent Robotics Industry (2026–2028)," targeting a "two zones, two centers" structure by 2028: the country's most dynamic cluster zone for intelligent-robotics innovation enterprises, a globally leading cluster zone for supporting hardware and software, plus an international robot-friendly demonstration city center and a Guangdong-Hong Kong-Macao Greater Bay Area robotics open-source ecosystem center. Named technology priorities include autonomous embodied-AI toolchains, embodied-AI models, and self-sufficiency in humanoid robot hardware and core components. The investment-attraction targets are more specific, aiming to draw companies in three categories — embodied-AI foundation models, dexterous hands, and planetary roller screws — to set up in Shenzhen. The plan also calls for promoting the use of Chinese-developed operating systems in robotics, porting over core ROS toolkit functionality, and using open-bidding mechanisms to encourage companies to start from lightweight use cases.Source: Securities Daily source

Altman: OpenAI will definitely develop humanoid robots · humanoid ⚠️ Remarks from an interview

"We will definitely develop humanoid robots, and we'll also make robots in other forms." OpenAI CEO Sam Altman said this on the Sources Podcast. His reasoning is that the real world is largely designed around the human body — computers and cars are both built around people — but he immediately added: the robot's form isn't the most important thing; what really matters is the robot's brain. Where a humanoid form offers no practical advantage, OpenAI will design non-anthropomorphic machines for specific users. In the near term he pointed to industrial settings like data center construction and maintenance rather than ordinary homes, with general-purpose robots for individual consumers remaining a long-term goal. OpenAI formed a robotics division this year, with job postings covering custom electromechanical actuators, simulation pipelines, and robot data collection.Source: IT Home source

USDOT releases national autonomous-vehicle strategy for fiscal years 2026–2030 · autonomy

"America Leads" lines up commercial trucking regulations, safety standards, emergency response, and cross-state data coordination all onto the federal agenda, covering through fiscal year 2030, building on NHTSA's April 2025 framework. The item with the biggest impact on freight is a rulemaking underway at FMCSA, which would separate the parts of existing commercial-vehicle regulations that assume an occupied cab from the parts that are truly necessary regardless. NHTSA is separately preparing an ANPRM (advance notice of proposed rulemaking) to set objective performance standards for automated driving system capabilities; USDOT calls this the first safety standard addressing autonomous-driving capability, while also stating plainly that it remains an early step. The division of authority stays the same: NHTSA regulates automated driving systems as motor vehicle equipment, FMCSA oversees the operation and safety of trucks and buses equipped with such systems, and states and localities retain their traditional authority over licensing, enforcement, and right-of-way rules. Specific requirements await the proposed rule and public comment process.Source: ACT News source

ICE plans to procure robot dogs, drawing opposition from privacy groups · adjacent

A procurement planning document from last month states that Immigration and Customs Enforcement intends to spend $1 million to $2 million on Boston Dynamics' Spot units, citing improved situational awareness and hazard assessment; this week it issued a further market-research notice inquiring about "quadruped unmanned ground vehicles." Will Owen, communications director at the Surveillance Technology Oversight Project, put it this way: "The last thing ICE needs is a camera-equipped robot dog that can open doors." A Boston Dynamics spokesperson responded that Spot's typical uses are toxic-gas detection, unexploded-ordnance inspection, suspicious-package screening, search and rescue, and post-disaster structural assessment, stressing that any attempt at weaponization is explicitly barred by the company's terms, ethical guidelines, and the anti-weaponization open letter it led along with five industry peers. An ICE spokesperson said the technology would not replace law enforcement personnel or their judgment.Source: FedScoop source

Xuanchuang (Chinese industrial robotics maker) unveils new-generation explosion-proof wheeled inspection robot · embodied

The unit has obtained dual explosion-proof certifications — Ex db eb mb ib ⅡC T6 Gb and Ex tb ib ⅢC T80℃ Db — plus IP66 ingress protection; its 580mm-narrow body fits through 700mm plant corridors, and its four-wheel, eight-drive omnidirectional chassis supports lateral movement and in-place turning. It runs 157 TOPS of on-device compute in parallel to handle multimodal inference for tasks like smart meter reading, equipment temperature sensing, and leak detection, with 8 hours of full-charge battery life and support for autonomous recharging. Its proprietary XBOTAEGIS architecture has three layers: the top layer understands tasks and makes decisions, the middle layer verifies inspection paths and action compliance, and a bottom-layer safety barrier resides permanently on the unit, locking the brakes within milliseconds if it identifies a risky command. Following an order of over 100 units from PetroChina (previously reported), the company has now deployed this system at gas storage facilities and chlor-alkali plants, where equipment-area corridors leave less than 5cm of clearance on either side. Founder and CEO Fu Zhe said: "Explosion-proof certification is just the entry ticket — the real difficulty is understanding operational workflows and safety standards." The company has already deployed a VLA execution model and is iterating tests based on an open-source JEPA world model.Source: LieYunwang Select source

Hardware · Supply Chain

· Qualcomm Dragonwing Q-2390 / IQ-2390: released September 1 and shown at IFA this week; the 1.1 TOPS IQ2 series sits at the entry tier of Qualcomm's IoT product line, with a quad-core Kryo CPU and Adreno 704 GPU; some SKUs offer an optional 600MHz SiFive E61 RISC-V real-time core, letting Zephyr handle sub-millisecond deterministic motion control while the Arm cores handle AI and connectivity, consolidating what previously required two separate chips into one source

· Uniview robot vision modules: Uniview's board secretary told investors that Uniview has partnered with DEEP Robotics and Xiaoyuan Robotics to ship vision modules in volume for inspection robots, positioning itself as a vision-component supplier rather than building complete robots in-house; the company has not given a revenue-share outlook for this business source

· TI CAN XL transceiver: Texas Instruments says it has released the industry's first commercial CAN XL transceiver, targeting robotics applications source

FutureX · Physical AI Daily — Issue 110 (09/05)

Shawn — Fri, 04 Sep 2026 15:02:23 +0000

Today's Highlights

· NHTSA opens an investigation hours after Cybercab hits the road, probing Tesla's basis for self-certifying compliance

· Draft revision of China's Road Traffic Safety Law adds a dedicated autonomous driving chapter, putting liability for violations while the function is activated on automakers

· Huandong Technology (RV reducer maker, subsidiary of Shuanghuan Transmission) withdraws its STAR Market IPO at the last moment, taking China's top RV reducer maker off the capital markets for now

· UBTECH lands a 1.5 billion yuan project in Shanxi, sending humanoid robots underground into coal mines

· Wandercraft signs 12 Calvin-40 customers, with Renault planning to deploy 350 units within 18 months

· Axis open-sources 50,000 Franka simulation trajectories, lifting π0.5's score on LIBERO-Plus by 4.9 points

· Stardust Intelligence's SmoothRL raises real-robot dynamic throwing success rate from 39% to 94%

Research Progress

1,500 hours of bimanual household demonstration data open-sourced, used to train VLA model XR-2 · vla

Public data for bimanual household manipulation has long been stuck below a thousand hours; this release provides 1,500 hours of demonstrations in one go, covering everyday household tasks, and uses this corpus to train the VLA model XR-2. The authors run scaling experiments along two axes: adding more expert demonstrations, and using DAgger-style correction data generated through real-time human intervention for post-training. Success rate climbs steadily with data volume along both axes, with no inflection point observed within the range explored. The dataset is released alongside the paper.

Jiafeng Xu et al. · arXiv 2609.03591 source

BRIDGE: putting humanoid morphology design and whole-body control into a single optimization framework · locomotion

Designing hardware separately from whole-body control is a longstanding root cause of humanoid robots "moving unlike humans." This paper proposes a data-driven framework for co-optimizing morphology and control, along with a new metric that jointly measures kinematic retargeting fidelity to human motion and dynamic tracking performance. Against three baseline humanoids — Bumi, K1, and ToddlerBot — it achieves SOTA on every metric. The result is an open-source 88cm-tall platform called Bridge, with the control policy released alongside it.

Jianren Wang et al. · arXiv 2609.03497 source

MINERVA: 0.54M parameters is enough to hit 95.1% on LIBERO · benchmark

The LIBERO leaderboard is now dominated by billion-parameter-class VLAs; this paper instead asks how much capacity the benchmark itself actually requires. A policy with just 0.54M parameters runs 2,000 rollouts across four standard suites, achieving an average success rate of 95.1% — only 2.4 points below the π0.5 result reported by LeRobot, with 7,700 times fewer parameters and only 5–9 milliseconds per chunk on a laptop CPU. A more notable finding: a task-ID permutation probe that only alters the task-index mapping drops success rate to near-random, suggesting that standard LIBERO's instruction conditioning is largely memorizing tasks. On the LIBERO-Plus perturbation set, performance slides to 46%–56%.

Kohei Sendai et al. · arXiv 2609.03715 source

FailBench: using VLMs as judges of robot task success or failure — even the best scores only 0.77 · benchmark

Using VLMs to judge whether a robot task succeeded is becoming the default approach, yet cross-domain reliability has rarely been measured. FailBench collects 2,197 manipulation attempts from 14 public sources, 75% of whose failures occurred naturally; testing 13 VLM detectors, the best achieves an average balanced accuracy of only 0.77. Models specifically fine-tuned for failure detection actually score consistently lower than general-purpose VLMs and their own pretrained baselines. On contact-rich assembly tasks, scores fall below 0.60, and error analysis further finds that when evidence is ambiguous, models systematically bias toward judging "success."

Zaruhi Navasardyan et al. · arXiv 2609.03611 source

EGR: teaching VLAs which sensor channels actually matter · vla

Trained on limited and homogeneous demonstrations, VLAs readily learn spurious correlations between sensor modalities — a failure mode the authors call modality entanglement. EGR uses a frame-by-frame, sensor-by-sensor task-relevance signal for gating, adds two consistency constraints during training, and incurs zero extra cost at inference. In real-world tests with a dual-arm Kinova plus physical distractors, success rate rose from 30% to 85%; on a single-arm platform combining vision with GelSight tactile sensing, it rose from 55% to 70%.

Yue Yang et al. · arXiv 2609.03142 source

WISE: scheduling a world model's imagination cuts GPU time by 80% · world-model

World models let policies try and fail in imagination, but this is compute-expensive, and rollouts that run too long accumulate prediction error and feed back unreliable supervision signals. WISE's approach is to schedule imagination, triggering it only in states relevant to interaction, running bounded multi-view rollouts, and evaluating candidate futures using progress and completion signals. It improves both π0 and π0.5, cutting GPU compute time by roughly 80% compared to full-scale imagination.

Chenhao Zhang et al. · arXiv 2609.03681 source

RoboTok: turning internet human videos into searchable robot supervision · manipulation

Robot data is expensive, and long-tail tasks are especially hard to cover. Given a query video of human manipulation, RoboTok retrieves relevant demonstrations from internet video to train dexterous-hand policies. The method learns an implicit motion space from 3D hand trajectories in a human-centric reference frame, making manipulation behaviors comparable across different camera views, scenes, and occlusions, while remaining compact enough to support continuous internet-scale indexing. Both retrieval relevance and downstream task success rate outperform existing robot data retrieval methods.

Howard Qian et al. · arXiv 2609.03199 source

Raw data may stay local, but household information still leaks through what's exported · perception

Home robots keep raw sensor data on-device, but the structured representations exported to planners, cloud services, and logs can still leak household information. The authors quantify this across 120 AI2-THOR scenes: three navigation-export representations all achieve a task success rate of 1.000 and an average path ratio of 0.898, yet their linkability at the representation layer ranges all the way from 0.532 to 0.970. Replacing explicit goal labels with goal regions drops target-category macro-F1 from 1.000 to 0.077 while success rate remains at 0.995; geometric coarsening, meanwhile, comes at the cost of collision-detection utility. Neither dropping fields nor adding abstraction yields a general-purpose privacy ranking — privacy must be evaluated task by task.

Yuqiao Xu, Erman Ayday · arXiv 2609.03055 source

Other papers today: FWBC-VLA adds contact awareness to wheeled-legged robot VLAs via sensorless residual torque estimation, validated on whiteboard wiping and door opening (arXiv 2609.03889 source); R2S-Eval replaces repeated real-robot trials with real-to-sim calibration plus VLM preference scoring (arXiv 2609.03276 source); SV-WAM builds a surround-view world-action model into end-to-end driving while controlling inference overhead (arXiv 2609.03602 source); Drive-HWM guides driving decisions using dynamic latent variables from a hierarchical world model (arXiv 2609.03572 source); LaPla connects VLM discrete reasoning to continuous control via latent-space alignment planning (arXiv 2609.04070 source); MulDP enables quadrupeds to autonomously parkour-navigate complex terrain (arXiv 2609.03984 source); one paper re-examines world models in safety-critical embodied systems, arguing that high prediction likelihood and visual fidelity do not guarantee that evidence needed for safe decisions has been preserved (arXiv 2609.03774 source); and another offers a unified survey of robot learning spanning representation learning, VLA, and world models (arXiv 2609.03927 source).

Open Source · Tools · Benchmarks

· Axis Sim Dataset V1: a Franka simulation manipulation dataset open-sourced by Axis Robotics, comprising over 50,000 human teleoperation trajectories, 207 manipulation tasks, and over 60,000 scene variants, using a simulated Franka Research 3 as the embodiment. Continued pretraining of π0.5 on this data raised LIBERO-Plus success rate from 83.9% to 88.8%, outperforming a comparably sized RoboCasa365 baseline by 37.3%; the company says performance kept improving as data scale grew from 25% to 100%. The dataset was co-built with researchers from UC Berkeley, Johns Hopkins, and the University of Michigan, with V2 planned to expand to 1.2 million trajectories and 1,200 tasks. source

· HandEdit: a dataset and evaluation benchmark open-sourced jointly by AgiBot (Chinese embodied-AI robotics company) with Shanghai Jiao Tong University and Fudan University, for the task of replacing human hands in first-person video with a specified robot embodiment while preserving object state, contact relationships, and background. It spans over 200 million image-editing samples and over 300,000 video clips, drawing on five egocentric datasets — EgoDex, ARCTIC, OakInk2, HOI4D, and HO-Cap — covering 26 URDF embodiments, over 600 scenes, and over 1,100 objects. The benchmark uniformly tests 11 editing methods including GPT-Image, Nano Banana, FLUX, and Qwen-Image-Edit, and one finding is that high VLM-assigned scores do not track with whether the interaction is actually correct. source

· SmoothRL: an online reinforcement learning framework released by Stardust Intelligence (Chinese robotics startup)'s foundation model team, with a technical report published on its website. Large-model inference is slow, and real-world deployment generally relies on asynchronous inference, so the actions a model "planned" and the actions the robot "actually executed" end up mismatched; SmoothRL splits an action chunk into committed, executing, and discarded regions based on execution state, computing gradients only over the executing region. Using a π0.5 policy fine-tuned per task on the cable-driven Stardust S1 embodiment, dynamic throwing success rate rose from 39% to 94% (250 rollout episodes), placing a pen cap rose from 8% to 83%, and unboxing a package rose from 30% to 90%. source

Funding and Deals

Huandong Technology (RV reducer subsidiary of Shuanghuan Transmission) | STAR Market IPO | Spinoff terminated, all filings withdrawn · hardware

On the evening of September 3, Shuanghuan Transmission announced it was terminating the spinoff listing of Huandong Technology and withdrawing all application documents, citing "significant changes in market conditions since the plan was first drawn up." Counting from the application's acceptance in November 2024, the push toward listing lasted under ten months. Huandong makes RV reducers, the core component in the major hip, shoulder, and knee joints of industrial robots; its share of the Chinese market reached 24.98% in 2024, up from just 10.11% four years earlier, over the same period biting Nabtesco's Chinese market share down from 51.77% to just over 40%. The sharpest question across two rounds of regulatory inquiry concerned customer concentration: sales share from the top five customers rose from 79.45% to 92.12% during the reporting period, with Estun, Efort, Inovance, and Siasun accounting for nearly all of its revenue, while its largest customer, Estun, swung to a loss of over 800 million yuan in the same period. The company's answer was that it had cut prices to gain share — gross margin fell from 42.47% to 35.36% in 2024, in exchange for market share jumping from 18.89% to 24.98%; in 2025, revenue was 438 million yuan, non-GAAP net profit was 75 million yuan, and gross margin recovered to 36.57%. The withdrawal doesn't change demand-side fundamentals: of the 10 to 14 joints in a humanoid robot, four to six heavy-load joints depend on RV reducers.Source: Phoenix Finance source

XTEND | SPAC merger listing (NYSE: XTND) | $1.5 billion valuation · adjacent

Israeli defense robotics company XTEND completed its merger with JFB Construction Holdings, renaming itself XTEND AI Robotics and beginning trading on the NYSE on September 4, with the merger valuing XTEND at $1.5 billion. The company was founded in 2018 by CEO Aviv Shapira along with three other co-founders, building human-operated, machine-executed tactical drone systems. Its list of strategic investors includes Eric Trump, Protego Ventures, Unusual Machines, and Tel Aviv University's TAU Ventures.Source: Globes

Weifen Zhifei (Hangzhou) | Series A2 | Hundreds of millions of yuan · adjacent

Led by Puhua Capital, with Hongxin Fund, Wuyuan Capital, Shenzhen Capital Group, Hongtai Fund, and BV Baidu Ventures participating. Founded in 2024 and led by Gao Fei, a doctoral supervisor at Zhejiang University, the company works on embodied aerial intelligence, aiming to let aircraft navigate autonomously and complete tasks in unknown environments with no GPS, no communications, and no lighting. Mining is where it has been most proven in practice — its P300 product replaced what would have been over two years of work at a large Chinese mining company in six months, covering more than 300 mining sites; it is now shipping in small batches with hundreds of units on order. Reporting on the funding round differs: JW Insights records it as Series A2, while Sohu's same-day funding roundup records it as Series A+.Source: JW Insights source

Hivebotics (Singapore) | Series A | $6 million · adjacent

Led by Vertex Ventures Southeast Asia and India, with participation from Fareast Land Development and bathroom fixture manufacturer Rigel. Its robot, Abluo, is fitted with a multi-jointed robotic arm to clean toilets, urinals, sinks, and floors in commercial restrooms, addressing what the company says existing commercial cleaning robots can only do for floors. Over the past 12 months, Abluo has logged roughly 10,000 operating hours across 20 sites in Asia, Europe, the Middle East, and North America, including hospitals, airport terminals, and shopping malls, with the company saying it replaces 30 minutes of manual labor with a five-minute human inspection. Co-founder and CEO Rishab Patwari said the company started with restrooms because they are the hardest setting — wet, cramped, and full of moving parts.Source: The Business Times source

Shenpu Intelligence | Pre-A+ round | Hundreds of millions of yuan · embodied

Announced complete on September 3. Funded by Kexi Capital, an intellectual property fund under Shanghai Science and Technology Innovation, Goldenpool Investment, Bozhou Industrial Investment, Chengshang Capital, and Qianrong Capital, with existing investor Linear Capital following on. The company positions itself as an AI-native embodied intelligence systems provider, building its system around four pillars: models, data, embodiment, and scenarios.Source: Guandian.cn source

Commercialization and Deployment

Cybercab begins carrying passengers in Austin, NHTSA opens investigation hours later · autonomy

NHTSA announced an investigation into Tesla on Friday morning, just hours after the first Cybercabs hit the streets of Austin. The vehicle has no steering wheel, no pedals, and no conventional rearview mirrors — none of the manual controls required under Federal Motor Vehicle Safety Standards remain. Tesla told regulators that, as is customary, it self-certified the vehicle's compliance with all FMVSS. NHTSA's filing states it is investigating "the processes and technical data Tesla relied on when certifying the Cybercab," as well as the extent to which the company determined that certain federal standards do not apply to this vehicle. NHTSA Administrator Jonathan Morrison left room for both sides: "NHTSA fully supports the safe development and deployment of autonomous vehicles. But as the federal regulator, we need to ensure all laws are being followed." Zoox went down this same road before — it self-certified compliance in 2022, was hit with a special order the following year triggering the same kind of audit inquiry, saw its commercialization timeline slowed, and eventually completed a Part 555 temporary exemption process, receiving final approval only in July 2026. Fleet size is another variable: according to Texas Department of Motor Vehicles data, Tesla newly registered 38 Cybercabs for its local robotaxi fleet as of August 31, bringing the total to 45 units. The vehicle relies on 8 HD cameras paired with an end-to-end neural network, with no lidar and no high-precision maps, using the same AI4 compute platform found in the Model 3 and Model Y. Tesla says that at scale, cost per mile could fall to $0.20. In China, plans remain limited to static display — starting mid-September, the vehicle will be shown in Beijing, Shanghai, and other cities, with no sales involved and no indication of plans to launch a robotaxi service in China.Source: TechCrunch source, CnEVPost source

Ruigan Robotics obtains China's first dual gas-and-dust explosion-proof certification for a wheel-legged quadruped · industrial

The certificate, issued by the Shanghai Instrumentation & Automation Systems Inspection & Testing Institute, carries the rating Ex IIC T6 Gb & Ex IIIC T80℃ Db, covering both gas and dust explosive environments. IIC corresponds to the highest-risk gas category, including hydrogen and acetylene, with T6 requiring the equipment's maximum surface temperature not exceed 85°C; IIIC covers conductive dust, capping surface temperature at 80°C. Beyond the whole-machine certification, the company also obtained independent certification for its core explosion-proof components. The significance lies in the threshold this clears — in petrochemical, oil and gas, and coal mining settings, most robots to date have carried only single-gas explosion-proof or lower-tier certifications, locking them out of these environments. The product behind the certification is the Delta explosion-proof wheel-legged inspection robot, which supports 50kg of top-mounted payload, switches between wheeled and legged modes to cross ditches and steps, and carries onboard lidar, an infrared thermal camera, HD cameras, sonar, and gas detectors, paired with the Robo-AI petrochemical cloud platform for anomaly recognition.Source: iFeng Tech source

PaXini brings tactile sensing onto BYD's final assembly line · embodied ⚠️ Vendor claim

The two parties signed a deep strategic cooperation agreement bringing PaXini (Chinese tactile-sensing startup)'s multidimensional tactile sensing into manually operated key stations on BYD's final assembly line. BYD Group Vice President Luo Zhongliang and Board Secretary Li Qian attended alongside PaXini founder and CEO Xu Jincheng, with the agreement signed by Zhao Weibing, director of BYD's Future Lab, and Luo Xiaoheng, PaXini's Chief Strategy Officer. The press release explains why final assembly is the sticking point: stamping, welding, and painting handle model changeovers with flexible fixtures and vision-based positioning, switching over within two hours; but final assembly's hundreds of thousands of clip, adhesive, connector, and fitting operations require recalibration for every new model and still rely on veteran workers' feel for material hardness and deformation variance. At a takt time of one vehicle every 1.4 seconds, these operations have so far resisted being absorbed by vision-based systems. The release gives no numbers on units installed, stations covered, or any acceptance metrics — this is a cooperation signing, not a deployment milestone. Both parties say they will jointly advance the Ministry of Industry and Information Technology and State-owned Assets Supervision and Administration Commission's special program for real-world humanoid robot and embodied intelligence training.Source: China.com source

Realman's RealBOT stationed at Beijing Daoxiangcun to make mooncakes · embodied

Ahead of the Mid-Autumn Festival, Realman Robotics' wheeled humanoid robot RealBOT was stationed at the No. 0 treasure-hunt outlet of Beijing Daoxiangcun (well-known Beijing pastry brand), working alongside store staff to make Beijing-style mooncakes, with the in-store deployment running through the Mid-Autumn Festival. This is not their first collaboration — at WRC 2026, the two had already demonstrated the process of making Beijing-style "flaky pastry" five-kernel mooncakes using the GLN remote operation network. The explicit purpose of this round is to record pouring angle, mixing force, and unmolding timing as real-robot data to feed back into the model. Founded in 2018, the company self-develops all four of its core components in-house; its joint modules currently have a maximum annual production capacity of 500,000 units, with a target of reaching a million units annually by 2026.Source: Gasgoo source

Wandercraft signs 12 customers for Calvin-40, Renault plans to deploy 350 units in 18 months · humanoid ⚠️ Vendor claim

Calvin-40 is this French company's humanoid robot for material handling, with customers spanning automotive, logistics, retail, healthcare, and defense; the customer list has not been disclosed. The only publicly named customer is Renault Group, which plans to deploy 350 units within 18 months. Both the 12 customers and the 350-unit deployment figure come from the company's own press release and have not been independently confirmed.Source: GlobeNewswire source

Industry Developments

Draft revision of Road Traffic Safety Law adds dedicated autonomous driving chapter, dividing liability by "whether the function was activated" · autonomy

A draft revision of China's Road Traffic Safety Law has been submitted for first reading to the Standing Committee of the National People's Congress and is now open for public comment; it adds a new dedicated chapter, "Special Provisions for Autonomous Vehicles." This marks the first time China has clarified the legal status of autonomous vehicles, the conditions for operating them on public roads, and liability for violations at the national law level. Previously, relevant rules existed only in departmental regulations, normative documents, and national standards; while some localities had attempted their own legislation, its scope was limited and standards for allocating accident liability were inconsistent. The draft defines an autonomous vehicle as one that continuously performs the entire dynamic driving task in place of a human driver, within its designated operational design domain. Liability is allocated accordingly: for road traffic safety violations occurring while the autonomous driving function is activated, the manufacturer or importer bears responsibility for handling the matter; if the company believes the violation is unrelated to the autonomous driving function, the burden of proof falls on the company. Vehicles with the function deactivated, and vehicles with only assisted-driving functions, continue to be regulated as non-autonomous vehicles. The insurance provisions set only a framework — mandatory traffic insurance will apply to autonomous vehicles, with specific implementation rules to be set by the State Council. The draft also closes a marketing loophole, requiring companies to accurately disclose the vehicle's operational design domain and prohibiting product descriptions or commercial promotion that misrepresent it. Two scholars interviewed offered differing assessments. Zheng Fei, a professor at China University of Political Science and Law, argues that the legal nature of the "human-vehicle handover" remains undefined — neither takeover duration nor operating standards have been set, making liability for accidents during a handover still hard to determine, and civil compensation, product liability, and criminal liability will all need to be spelled out through supporting regulations. Zhang Li, deputy dean of the law school at the same university, takes a different view, arguing that the Road Traffic Safety Law is not currently the right vehicle for regulating civil liability at all — doing so properly would require adding manufacturers, operators, and system developers as liable parties alongside humans, and that "the timing may not yet be ripe."Source: Yicai source

UBTECH lands 1.5 billion yuan project in Shanxi, sending humanoid robots into coal mines · humanoid

On September 2, the Shanxi Comprehensive Reform Demonstration Zone signed an agreement with UBTECH and Beijing Hehao Ruisheng for the "Shanxi Embodied Intelligence Innovation Center and Industrial Energy Mining Special Robot Intelligent Manufacturing Project," with total investment of 1.5 billion yuan, to be built at Xiaohe New Industry Park. The project has two phases: the first is a data collection and scenario testing center, and the second is a UBTECH robot factory, aiming to become a core intelligent manufacturing base for robotics in North China. The choice of Shanxi comes down to coal. Shanxi's above-scale raw coal output was 1.305 billion tonnes in 2025, roughly 27% of the national total of 4.83 billion tonnes and the highest among all provinces. According to the Shanxi Provincial Energy Bureau in December 2025, the province had built 275 intelligent coal mines and 1,594 intelligent extraction faces, with intelligent mines accounting for 63% of production capacity, and underground substations and pump rooms now largely unmanned. What remains are inspection, maintenance, and auxiliary transport roles — high-risk jobs that still require mobility, exactly the segment special-purpose robots are meant to take on. Underground environments demand a level of explosion-proofing, dust resistance, reliability, and remote control beyond what a factory floor requires, making the case for paying for robots easier to argue than "humanoids on car assembly lines." Skepticism persists. According to an April 2025 report by Xinhua-affiliated Economic Information Daily, Boston Dynamics founder Marc Raibert has said that car manufacturing is already a strength of traditional industrial robots, and that humanoid robots can play only a limited role on the production line, more like a "mascot"; Abderrahmane Kheddar, a member of France's National Institute for Digital Science and Technology and professor at the University of Montpellier, likewise believes humanoid robots' advantages have yet to be fully demonstrated.Source: OFweek Smart Manufacturing (republished by Eastmoney) source

Inside a Beijing humanoid data facility: real-robot data runs 500–1,000 yuan per effective hour, and prices are still falling · adjacent

A reporter from STAR Market Daily visited the data and training facility of the Beijing Innovation Center for Humanoid Robotics. Spanning over 6,000 square meters in total, with roughly 5,000 square meters on the first two floors dedicated to data collection and training, the facility has built out more than thirty realistic scenarios across six major categories — home, retail, industrial, healthcare, and others — and has deployed over 150 sets of embodied robot equipment plus more than 150 sets of non-embodied capture equipment. The operating rhythm resembles a factory: collectors work 8-hour shifts, and after accounting for setting up scenes and props, get about 6 effective hours of recording; data is transmitted as it's collected, processed overnight by an automated pipeline, and delivered the next day — an internal cycle nicknamed "7+1." Annual capacity is 180,000 hours, and the facility has served over 100 clients. The most concrete figure in the piece is pricing: real-robot data costs roughly 500 to 1,000 yuan per effective hour and has already come down, while non-embodied data ranges from tens to one or two hundred yuan; the facility acknowledges it does not currently make money selling data. The person in charge pushed back directly on industry claims: "Some organizations say they'll reach ten million hours by year-end; others say they already have tens of millions of hours of non-embodied data — but how much of that is genuinely high-quality data?" Their standard is data that produces real improvement in a given scenario, likened to the truly valuable "corner cases" in autonomous driving. Under a three-stage quality check, the data delivered to one leading client in the first half of this year passed validation at over 95%. On the open-source side, the center's open-sourced RoboMIND dataset has recently surpassed 20 million downloads worldwide.Source: Cailianshe · STAR Market Daily source

XPeng's second-generation VLA 6.3.0 rolling out this month, model now retains 30 seconds of memory · autonomy

The new version begins rolling out to all Ultra and Ultra SE models starting in September, launching first on the G9L Ultra and Ultra SE, with Max models running a single Turing chip receiving a distilled version of the second-generation VLA, VLA Lite, the same month. Three modules correspond to three time horizons: Infini-VLA sets historical memory in the production version to 30 seconds, X-Foresight projects 6 seconds ahead, and Streaming Inference in between takes in input, runs inference, and outputs trajectories continuously, which XPeng says improves decision speed by 300%. Model parameter count has grown 3.5x, and data throughput per training run has reached 100 million video clips. The claim of "over 20x improvement in overall safety capability" comes from XPeng's own simulation and testing methodology.⚠️ Vendor claim The same architecture is also being ported to robotaxi use — internal testing has completed over 2,000 orders, and a vehicle equipped with the second-generation VLA recently obtained a remote-testing qualification for intelligent connected vehicles from Guangzhou authorities, allowing driverless-in-driver-seat testing on designated roads.Source: IFanr source

Yangtze River Delta G60 cities issue joint initiative on embodied intelligence quality standards · industrial

At the 9th Yangtze River Delta G60 Sci-Tech Innovation Corridor Quality Standards Conference in Hangzhou on September 4, participating cities released the "Yangtze River Delta G60 Cities Joint Initiative on Embodied Intelligence Industry Quality Standards," promoting standards across foundational/general technology, core technology, components, complete systems, and industry applications, and proposing cross-city collaborative governance of data. Hangzhou provided much of the basis for this initiative, having clustered over 700 robotics-related companies, with the local industry cluster's output value surpassing 106.8 billion yuan in 2025, and its market share in China for quadruped robots and humanoid robots exceeding 80% and 50% respectively. There has already been movement on standards: China's first group standard for maturity evaluation of quality management systems in the humanoid robot industry, T/ZRIA 002—2026, was initiated in November 2025 and took effect on June 19, 2026.Source: CNR source

CICC: China's industrial robot density has gradually surpassed that of the US, Japan, and Germany · industrial

CICC (China International Capital Corporation) Research released "The Impact of Factor-of-Production Shifts under New Quality Productive Forces on the Economy and Markets" on September 4, observing that China's industrial robot density has risen rapidly over the past decade or so and has in recent years gradually surpassed that of the United States, Japan, and Germany, though the report does not provide a specific density figure. Accompanying this trend, the synchrony between fixed-asset capital investment and new hiring has slowed at some manufacturing firms, with labor freed up from manufacturing flowing increasingly into services.Source: Guandian.cn source

Berg Insight: humanoid robot annual shipments to reach 26 million units by 2040 · humanoid ⚠️ Projected figures

This research firm projects a path from 16,000 units in 2026 to 26 million units in 2040, a 63.7% compound annual growth rate, corresponding to a market size growing from $890 million to $554 billion. In the same forecast, humanoid-robot-related cellular connections grow from 5,000 in 2025 to 50.6 million in 2040. The firm counts over 100 companies currently developing full-size humanoids, most still at the research, prototype, or pilot-deployment stage.Source: IoT Business News source

Hardware · Supply Chain

· Dasheng-1 reconfigurable dexterous hand: released September 4 by Heimai Technology (Chinese robotics component maker) jointly with LimX Dynamics, with 4 fingers and 11 active degrees of freedom, single-hand grip payload ≥24kg, repeat positioning accuracy ≤0.03mm, fingertip force ≥18N, and control frequency above 100Hz; the maker says it can switch between 7+ working configurations with zero downtime, replacing multiple gripper sets that would otherwise need to be swapped on a production line. source

· Black Sesame Technologies Aura module: Aubo Robotics (Chinese collaborative robot maker) has completed delivery of its first batch of Aura development kits in July, for use in collaborative robot development. Aura belongs to the SesameX platform, aimed at multi-legged robots and collaborative arms, offering 70 TOPS of heterogeneous compute; the platform spans three tiers from the 16 TOPS Kalos to the 700 TOPS Liora. source

· Uniview inspection vision modules: Q Technology (Group) says its subsidiary Uniview has partnered with DEEP Robotics and Weilan Robotics, with vision modules for inspection robots now shipping in volume. source

FutureX · Physical AI Daily — Issue 109 (09/04)

Shawn — Thu, 03 Sep 2026 16:56:02 +0000

Today's Highlights

· Uber and Wayve launched the UK's first paid robotaxi service in London, with no lidar and no geofencing

· Figure signed a $3.5 billion compute deal with Nscale, with deployment of up to 100,000 GPUs

· According to Bloomberg, Waymo raised its first-ever debt financing of $3 billion, backed by Pimco, Blackstone and Sixth Street

· Horizon Robotics' licensing and services revenue surpassed hardware for the first time in H1, reaching 55% of total revenue

· PlusAI (Chinese autonomous trucking startup) attempted a SPAC merger for the third time, at a pre-money valuation of $800 million — a quarter of its 2021 valuation

· CJ Logistics put two humanoid robots to work on packaging in an Olive Young warehouse

· UBTECH's U1 received 13,361 pre-orders, with first deliveries on September 16; the VP called mass production a huge challenge

Research Progress

A single human demonstration teaches a dexterous hand to grasp 16 objects · manipulation

Data collection for dexterous hands has long been the bottleneck for multi-finger manipulation. A Stanford team (including Fei-Fei Li and Dorsa Sadigh) changed the reward from "task completion" to "whether the local geometry near contact points is correct." The resulting single policy transferred on real hardware across 16 objects, 4 task types, and 2 hand types, achieving a 71% success rate with the smallest sim2real gap among compared methods. Shape, size, mass, and friction can all vary as long as the contact structure is preserved.

Satvik Sharma et al. (Stanford) · arXiv 2609.01938 source

Before an emergency stop, a humanoid first calculates whether it can actually stop · locomotion

Executing a fixed motion upon receiving an emergency stop command is standard practice for current humanoids, regardless of whether stopping safely is actually feasible in the current state. Pieter Abbeel, Koushil Sreenath, Guanya Shi and others reformulated emergency stopping as a reach-avoid problem, paired with two complementary stoppability estimators: one supervised by the actual outcomes of a fixed stopping policy, and one supervised by physical recoverability derived via Hamilton-Jacobi reachability. Both only stop when stopping is deemed feasible; otherwise the system switches to a damped fall policy. The estimators don't depend on the behavior policy used before stopping, so they don't need retraining for new tasks.

Junfeng Long et al. (UC Berkeley / CMU) · arXiv 2609.02358 source

DeepMind releases a multi-view 3D point tracking benchmark; none of 30+ methods come close to solving it · benchmark

Multi-camera setups are increasingly used in robotics, AR/VR, and autonomous driving, yet no benchmark has tested long-horizon 3D point tracking with moving cameras and synchronized multi-view feeds. TAPVid-MV fills this gap: 284 sequences, 1,142 calibrated camera streams, and 109,769 point trajectories spanning robotics, human activity, driving, and synthetic scenes, with every trajectory manually visually verified. Across 30+ baselines, none come close to solving the task, and existing multi-view point trackers are not consistently better than monocular methods. By comparing reconstruction and tracking on the same data, the authors identify geometry recovery as the accuracy bottleneck.

Skanda Koppula et al. (Google DeepMind) · arXiv 2609.01899 source

Giving world action models depth perception · world-model

World Action Models (WAM) repurpose pretrained video diffusion models into robot policies that jointly predict future frames and actions — but they all consume only RGB, discarding 3D information. SA-WAM uses a nonlinear encoding to compress unbounded depth signals into the value range expected by a frozen VAE tokenizer, injecting geometry without modifying the tokenizer. A single diffusion backbone jointly outputs actions, RGB, and depth, achieving SOTA on RoboCasa and LIBERO-Plus, with even larger gains on a real UR5 in randomized environments.

Javier Alejandro Lopetegui Gonzalez et al. (Inria) · arXiv 2609.02531 source

Stepping stones and gap crossing: world models take over humanoid foothold decisions · locomotion

On terrain like stepping stones, narrow ladders, and gap crossings, feasible footholds are sparse and discontinuous, leaving no margin for a single misstep. WM-LOCO jointly trains a recurrent world model with a PPO policy, requiring no explicit foothold annotations, allowing the policy to plan beyond terrain that's immediately visible. In simulation, on both gap-crossing and stepping-stone tasks, baseline methods completely fail while this method succeeds; real-robot results have not yet been reported.

Yuxi Liu et al. · arXiv 2609.02542 source

Cross-embodiment zero-shot transfer — first, define what "zero-shot" actually means · vla

The term "zero-shot cross-embodiment" in VLA papers is often used inconsistently. ZETA splits it into two cases: the target embodiment never appears anywhere in the training data, versus one seen only during pretraining. With this distinction in place, the paper conducts a controlled analysis comparing four factors — including state-action representation — for their impact on tabletop manipulation transfer.

Mi Yan et al. · arXiv 2609.02546 source

Dissecting GR00T N1.5's action decoder · vla

VLAs connect language and perception to action, but what representations actually drive behavior internally has remained unclear. LAVLA proposes a latent-space clustering analysis framework, applying it layer by layer to NVIDIA's GR00T N1.5, with a focus on its action decoder.

Theodor Wulff et al. (Manchester / Hamburg) · arXiv 2609.02634 source

Human-to-humanoid motion retargeting, without hand-crafted mappings · locomotion

Turning large-scale human motion data into reference trajectories for robots is bottlenecked by differences in morphology, degrees of freedom, and joint ranges, with the standard approach being hand-written mappings for each robot model. UMR instead learns dense point-cloud correspondence, eliminating manually designed human-robot correspondence tables, with the goal of directly converting large-scale human motion libraries into trainable robot data.

Hanyang Cao et al. · arXiv 2609.02134 source

Other papers today: DiffuSearch unifies trajectory generation and refinement under a single objective, significantly reducing collisions (arXiv 2609.02252 source); CrashDiffuser lets a VLM specify impact locations to synthesize safety-critical scenarios (arXiv 2609.02270 source); DCRL's recursive value learning raises the average score across OGBench's five hardest long-horizon tasks from 55 to 64 (arXiv 2609.02237 source); HINT lets a VLA infer true human intent from a coarse instruction (arXiv 2609.02653 source); MS-MEM uses pushing and grasping to actively improve shelf mapping while reducing scene disturbance (arXiv 2609.02493 source); contact-rich elder-care tasks lack physically consistent evaluation, which the authors address (arXiv 2609.02402 source); a quasi-direct-drive knee prosthesis matches commercial models in mass, height, peak torque, and noise (arXiv 2609.02003 source); Sim2Signal decomposes the sim-to-real gap in traffic signal control into observation, action, transition, and reward categories (arXiv 2609.01676 source); a deployment-readiness evaluation protocol for LiDAR semantic segmentation under coarse labels, adverse weather, and domain shift (arXiv 2609.02830 source); MV-dVRK, the first multi-view synchronized ex-vivo surgical reconstruction dataset (arXiv 2609.02717 source); a survey on autonomous driving's shift from text-based chain-of-thought to action-grounded reasoning (arXiv 2609.01659 source).

Funding & Deals

Figure | Strategic Partnership + Strategic Investment | $3.5 Billion Compute Commitment · humanoid

The compute order alone reaches $3.5 billion, with intended scale exceeding $6 billion. UK AI cloud provider Nscale signed a multi-year strategic partnership with Figure, with potential deployment of NVIDIA's Vera Rubin platform reaching up to 100,000 GPUs, with the first batch targeted for online in Barstow, Texas in the second half of 2027, supporting training and inference for Figure's Helix model series. Nscale also made a strategic investment in Figure, becoming both a shareholder and its preferred compute supplier. The agreement also leaves room for a reverse arrangement: the two companies will explore using humanoid robots to expand Nscale's own supply chain. A single compute commitment exceeding $3.5 billion from a humanoid robotics company is a scale previously seen only from foundation model companies.Source: Nscale source

PlusAI | SPAC Merger | Pre-money Equity Value of Approximately $800 Million · autonomy

Third time's the charm, after two prior failed attempts. PlusAI (Chinese autonomous trucking startup) reached a merger agreement with blank-check company Texas Ventures Acquisition III Corp, at a pre-money equity value of approximately $800 million. The deal could bring in up to $300 million in funding, including over $60 million in already-committed financing and about $236 million held in the SPAC's trust account. The company, headquartered in Santa Clara, California, focuses on SuperDrive, an L4 system for commercial trucks, targeting commercial deployment by 2027; its HyperFoundry software development platform has already generated $25 million in revenue, with full-year contracted revenue targeted at $40-50 million. It currently operates driverless freight routes in Texas with Ryder and International, and partners with global truck manufacturers on factory-installed production. Background: in May 2021, a merger agreement with HCIC V valued the company at about $3.3 billion, but was terminated that November; a second deal with Churchill Capital IX collapsed again in April 2026, officially attributed to "market conditions." Today's pre-money figure of $800 million is less than a quarter of the 2021 valuation.Source: Reuters source | 36Kr source

Waymo | First Debt Financing | Over $3 Billion · autonomy ⚠️ Unconfirmed report

Waymo, which has relied entirely on equity funding until now, is turning to debt markets. According to Bloomberg, the company is in talks with Pimco, Blackstone, and Sixth Street for its first debt financing of over $3 billion, with Goldman Sachs serving as arranger; the spread would be over 500 basis points above the benchmark rate and would carry no credit rating. Earlier this year the company raised $16 billion in equity at a $126 billion valuation. It currently completes over 500,000 paid rides per week across 14 US cities, with a year-end target of 20 cities and 1 million weekly rides.Source: Bloomberg source

UBTOTECH | Series A+ | Hundreds of Millions of RMB · humanoid

The round was led by Orient Securities Capital, with new investors Jinan Xianxing Investment Group and Zibo High-tech Zone Venture Capital participating, and angel-round investor Legend Star adding further investment. This Shandong-based company, founded in 2014, builds legged robots; its self-developed humanoid, "Walker Taishan" E1, won Shandong's delegation its first gold medal in the tug-of-war event at the 2nd World Humanoid Robot Games, and its quadruped products have already entered use cases including highways, telecommunications, power, tourism, and ports. Founder and chairman Fan Yong received his PhD from the Institute of Automation, Chinese Academy of Sciences in 2000, and the core team comes from Shandong University's Robotics Research Center and the Institute of Automation. The funding will go toward next-generation high-power-density integrated joint modules, industrial and consumer humanoid robots, and heavy-load, explosion-proof quadrupeds.Source: iThome source

First China humanoid robot ETF lists on Nasdaq · adjacent

Ticker CROB, issued by Defiance ETFs, tracks the Solactive China Humanoid Robotics Index, with constituents required to be headquartered in China and listed in Shanghai, Hong Kong, or Shenzhen. The top ten holdings — Green Harmonic Drive, Inovance Technology, Lingyi iTech, Sanhua Intelligent Controls, Megmeet, Hengli Hydraulics, Topgroup, Luxshare Precision, CATL, and Wolong Electric Drive — are exclusively component and manufacturing suppliers, with no full-robot makers included. Defiance CIO Sylvia Jablonski said China has already succeeded in EVs, solar, and batteries, and views humanoid robotics as the next stage of strategic positioning. Previously, Golden Rui Fund's KOID was the first US humanoid-themed ETF, which already includes Unitree Robotics.Source: China Fund News source

Commercial Deployment

Uber and Wayve launch the UK's first paid robotaxi service in London · autonomy

Waymo has been beaten to the punch. Uber launched a robotaxi service in London, becoming the first company to offer paid autonomous rides in the city, using technology from UK startup Wayve. The fleet uses Ford Mustang Mach-E vehicles, with Nissan Leafs to follow; currently about 15 vehicles are registered with Transport for London. To ride, users must opt in via app settings, and there's no guarantee of a match when requesting UberX, Uber Comfort, or Uber Electric. Initially, a safety operator remains in the driver's seat, licensed by Transport for London; Uber says it does not intervene in driving and will phase this out. The technical approach diverges from the US mainstream in two respects: no lidar and no geofencing. Wayve describes its AV2.0 as a single learned AI driver that doesn't rely on hand-coded rule stacks or high-definition maps; spokesperson Julianne McGoldrick said the service will eventually cover all of London except airports. This marks the first large-scale test of driverless operation outside the US and China.Source: The Verge source | The Guardian source

Zoox brings paid robotaxis to Las Vegas airport · autonomy

Starting September 3, Zoox became the first company approved to operate driverless vehicles at Harry Reid International Airport, with pickup and drop-off available at both Terminal 1 and Terminal 3: at T1, drop-off is at the departures level and pickup at level 0; at T3, both pickup and drop-off are at level 0. The route connects the airport with the Las Vegas Strip. The company only began charging fares on August 10, under the first commercial exemption issued by NHTSA specifically for "purpose-built robotaxis." Local rideshare drivers have already publicly voiced concerns about their income.Source: TechCrunch | Las Vegas Review-Journal

CJ Logistics puts humanoid robots into real warehouse operations · industrial

Two humanoid robots have been deployed at an Olive Young distribution center in Yangji, Yongin, doing packaging work — placing cushioning material into boxes. The company says this marks the first time it has deployed humanoids in an actual logistics workflow, having previously only conducted on-site testing at its Gunpo fulfillment center. Plans call for expanding into picking, sorting, inspection, and packaging, with the goal of having a single robot handle multiple processes. On the technical side, CJ built its own robot foundation model (RFM), combining visual and sensor information with simulation data to let robots determine how to move around unfamiliar products and workstations; the hardware comes from Robotis, gripper technology from Aidin Robotics, and the RFM was developed by RealWorld AI. Two units don't amount to meaningful production capacity, but it shifts the question of whether humanoids can work in spaces designed for humans from demonstration to a production line shipping goods daily.Source: The Korea Times source

EACON's driverless system deployment surpasses 1,500 electric mining trucks · industrial

Nearly doubling in under six months, from 800 units in March to over 1,500 today. Within Easy Control's (Chinese mining autonomy company) autonomous driving deployments, pure-electric mining trucks now account for about 42% of the fleet, hybrid diesel-electric 41%, methanol hybrid 16%, and conventional internal-combustion trucks under 1%. At China Energy's Zhundong open-pit coal mine, 120 pure-electric trucks run on the ORCASTRA system with 28 charging stations, with dispatch simultaneously calculating battery levels, predicting energy consumption, tracking charging-station availability, and coordinating with production schedules — vehicles navigate to charging stations on their own and return to the transport line once charged. At Zijin Mining's Julong copper mine, 60 Sany SKT105ED trucks operate routinely at high altitude, maintaining an operational rate above 90% under low air pressure, extreme cold, and freeze-thaw conditions with rain and snow. China Baowu's Taihe iron mine added 12 more trucks in March after over a year running its initial 6 units. Mining sites are currently one of the few areas in commercial driverless vehicles where electrification and autonomy have both been achieved at scale together.Source: International Mining source

UBTECH's U1 delivery test approaches, as the substance behind 13,361 pre-orders remains to be seen · humanoid

First deliveries are set for September 16, with the company aiming to fulfill all orders within the year. The U1, launched on June 30, is a hyper-realistic companion robot aimed at emotional companionship; the launch event announced cumulative pre-orders of 13,361 units across all channels. That figure comes with caveats: pre-orders require only a RMB 3,000 deposit, which is fully refundable during the presale period, leaving the eventual final-payment conversion rate unknown. Production capacity is another hurdle — as of the end of 2025, UBTECH Robotics' annualized production capacity for full-sized humanoids was about 6,000 units, while the U1 is an entirely new consumer model requiring dedicated production lines for its silicone skin and facial micro-servos. Vice President Jiao Jichao told reporters after the launch event bluntly: "Mass production at a scale of over ten thousand units is a huge challenge." He attributed the difficulty to low yield rates on lifelike silicone skin and supply chain constraints from the large number of non-standard micro-servos throughout the robot. JD.com's flagship store lists the U1 Pro as available after September 15, with customer service quoting a 60-day shipping window after ordering. Capital markets reacted with initial excitement followed by a pullback: on launch day the stock rose over 18% intraday before closing up 7.48% at HK$102.8; by the September 3 close it had fallen back to around HK$79, a pullback of over 20% from the peak. On August 26, the U1 won the "AI-Driven Healthcare Innovation Award" for Asia-Pacific at the MedTech World Awards 2026 in Hong Kong; around the same time, UBTECH also jointly launched a wellness-care edition of its Umi robot with Meditech.Source: Red Star Capital Bureau source

Zhuji Dynamics' TRON 2 ties rebar on a construction site · embodied

Two TRON 2 dual-arm robots, one holding a nail gun and the other a wooden board, worked together to assemble and secure formwork, then laid and tied rebar across multi-layer scaffolding, completing a scaled-down construction process for a concrete tilt-up panel. The work was carried out by physical AI company ZINOVA, in partnership with construction robotics company RIC Robotics. ZINOVA's approach centers on "tool intelligence" — building its technology stack around general-purpose tool grasping and tool-interaction perception, aiming to make the same system reusable across tools, tasks, and robot embodiments rather than customizing a robot for every trade. Construction was chosen as its first validation scenario, citing both a shortage of skilled labor and the highly non-standardized nature of construction environments. The TRON 2 platform supports switching among dual-arm, biped, and wheeled-biped configurations, and was reconfigured here for large-space, multi-process requirements. This remains a scaled-down proof of concept for now, not routine operation on an active construction site.Source: Gasgoo source | 21st Century Business Herald source

Industry Developments

Horizon Robotics' licensing revenue surpasses hardware for the first time in H1 · autonomy

The revenue mix has flipped. Horizon Robotics' H1 revenue reached RMB 2.055 billion, up 32.9% year-over-year, of which licensing and services revenue was RMB 1.129 billion, up 52.7% year-over-year and accounting for 55% of the total — surpassing, for the first time, the RMB 926 million from products and solutions (up 14.8% year-over-year). Overall gross margin held at 66%, with gross profit of RMB 1.356 billion. On market share, the company's share of the Chinese-brand ADAS market surpassed 50.0% for the first time; its share of the urban NOA chip market rose from 17.9% in 2025 to 22.8%, moving from third to second place; and it holds the top spot in the overall Chinese-brand smart driving chip market with 31.9% share. Capital markets have long viewed Horizon Robotics primarily as an automotive chip company, but this half-year report shows the higher-margin licensing business is now the larger half.Source: Caixin source

Altman confirms for the first time: OpenAI will build its own humanoid robot · humanoid ⚠️ Remarks in a podcast

In a podcast released local time on September 2, Sam Altman, when asked about hardware progress, said "we will definitely do humanoid robots." His reasoning: the physical world is built around the human body — keyboards, doors, heavy machinery, kitchens. He also said OpenAI will simultaneously develop specialized robots for data centers, with near-term focus remaining on industrial infrastructure and data center automation, as well as the underlying cognitive models. No specific model, timeline, or team size was given.Source: The Paper source

Beijing's humanoid data facility opens; RoboMIND downloads surpass 20 million · world-model

The Embodied Intelligence Data and Training Base of the Beijing Innovation Center for Humanoid Robotics opened today. Built in under a year, the facility spans more than 6,000 square meters, covering six major categories — home, retail, office, industrial, medical, and wellness — across more than 30 representative scenarios, with over 150 types of robot hardware and more than 100 embodiment-free data collection setups. It has an annual capacity of 180,000 hours of high-quality data and has already delivered over 30,000 hours to leading clients. Its open-source dataset RoboMIND doubled in global downloads within a month, surpassing 20 million. Facility director Xia Hualin said the most critical competitive edge in embodied intelligence going forward won't be the robot body or the algorithm, but the data flywheel: "I originally came from autonomous driving. The biggest factor behind how quickly autonomous driving was deployed wasn't the vehicle or the large-model algorithms — those were already mature. What was needed most was data." He defined the most valuable data as "feedback data" — for example, if a robot can wash dishes at Household A but fails at Household B, quickly feeding the failure data from Household B back into training and redeployment. The facility says AI-model-driven automated labeling has improved data production efficiency by over 50% and reduced overall costs by over 30%; in H1, the data acceptance pass rate for a leading client delivery exceeded 95%.Source: Sina Technology source

LG Electronics teams up with NC AI to take on South Korea's edge-AI chip humanoid track · humanoid

South Korea's Ministry of Trade, Industry and Energy, together with the Korea Evaluation Institute of Industrial Technology (KEIT), selected a consortium of NC AI and LG Electronics for the humanoid track of its "K-Edge AI Semiconductor Technology Development Project." The goal is to use domestically made edge-AI chips to power autonomous humanoids capable of performing a variety of tasks in complex everyday environments, bringing robot hardware, training data, AI models, and domestic chips together under a single development framework. NC AI's (Korean gaming company's AI unit) role covers initial training data for General Motion Tracking (GMT) and a human-to-robot motion retargeting pipeline; it plans to use its own motion capture and animation production technology, combined with digital twin environments, to mass-generate task scenarios rather than collecting them one by one in real-world settings.Source: The Elec source

Unitree's wholly owned subsidiary lands in Chengdu, with a "three bases, two centers" layout · humanoid

Chengdu Yuchen Technology Co., Ltd. completed its business registration on September 2. Unitree Robotics' Chengdu plan calls for three bases plus two centers, including a planned flagship global launch center on Chunxi Road. As previously reported, Unitree's market cap had halved from its first trading day by its 11th trading day since listing; new this week is this regional expansion, along with internal management controversy sparked by the company's employee expense-reimbursement policy.Source: Pandaily source | Sina Finance source

Tesla's Cybercab event set for Thursday, with very few details · autonomy ⚠️ Planned event, unconfirmed details

Reuters reports that Tesla will hold a Cybercab-related event on Thursday, with the company not disclosing specific content. Around the same time, reports have surfaced of Austin residents already riding in Cybercabs. Production status, capacity, and the scope of any regulatory permit remain without verifiable public figures.Source: Reuters source

Hardware & Supply Chain

· Wonday Bearings: The "first stock in Segment 920" on the Beijing Stock Exchange reported H1 revenue of RMB 201 million, up 12.24% year-over-year, of which robotics-related business revenue was about RMB 8.01 million; its higher-margin slewing bearing business grew 22.90% year-over-year with a 46.76% gross margin. The company has developed over 600 models across 11 major series of slewing bearings adapted for robots, cumulatively serving over 100 clients (including Hikvision Robotics, Sany Robotics, KUKA, and Siasun, mostly in sample validation or small-batch stages); crossed-roller bearings for harmonic reducers are about to ship for bench testing, and angular contact ball bearing samples for RV reducers have completed precision and lifespan validation. Chairman Xu Qunsheng put it plainly: high-end component R&D requires heavy investment and long certification cycles, so results won't materialize in the short term. source

· Multimodal robotic sensing: At SEMICON Taiwan 2026, physical AI was identified as a new wave of demand for advanced sensors, with robots moving from controlled demonstrations to real-world deployment driving orders for multimodal sensing (DIGITIMES). source

This Week's Watch

France launches Tesla FSD road testing, just two months after calling it "too risky" · autonomy

On September 2, Paris approved on-road testing of Tesla FSD in France, with Tesla providing two vehicles equipped with the system. French Transport Minister Philippe Tabarot said France would independently verify data provided by the Netherlands and Tesla, and test the system under French road conditions, with completion planned for mid-to-late September. This marks a clear reversal — in July, he had said the safety tradeoffs weren't yet sufficient to justify authorization, but he now describes "constructive exchanges" with Elon Musk, with the two sides having cooperated on technical adaptation for months. An EU-level approval vote is imminent, and France's test findings will feed directly into that process.Source: Euronews | Reuters

China's first systematic overhaul of its road traffic law gives autonomous driving its own dedicated chapter · autonomy

The draft revision comprises 9 chapters and 170 articles, submitted for first reading at the 24th session of the Standing Committee of the 14th National People's Congress on August 25, and opened for public comment starting August 28; within less than a week, over 8,000 people had submitted over 12,000 comments. The current law has been in effect since 2004, with only three prior amendments touching individual provisions. The new dedicated chapter on "Special Provisions for Autonomous Vehicles" establishes several firm rules: autonomous vehicles must pass a road traffic rules compliance test and be legally registered before operating on public roads; in the event of a traffic violation while autonomous driving functions are activated, the manufacturer or importer is held responsible, and if the company believes the violation is unrelated to the autonomous driving function, the burden of proof falls on the company itself; and traffic police authorities may retrieve relevant vehicle data in the event of an accident. NPC Standing Committee member Huang Ming said the purpose of the dedicated chapter is to "draw clear regulatory red lines for the new business model and clarify a compliant development path." The reversal of the burden of proof — previously accident attribution had long relied on manufacturers' own claims — has now been written into the draft law.Source: China News Service source

Goldman Sachs sharply raises humanoid shipment forecast: 890,000 units by 2030 · humanoid

In a research note published September 2, Goldman Sachs raised its forecast for global humanoid robot shipments in 2030 from a previous 256,000 units to about 890,000 units, and its 2035 forecast from 1.4 million to about 6.5 million units, corresponding to a market size of about $138 billion; this year's baseline shipment estimate was also raised from 51,000 to 75,000 units. On deployment sequencing, the bank ranks logistics and warehouse operations as the leading use case, with the automotive industry as the second wave. The nearly 3.5x upward revision compares against the same institution's 2035 market-size forecast of $38 billion issued just over a year ago.Source: China Review News

Arthur D. Little: humanoids account for only $2-3 billion of an $18 billion physical AI market · adjacent

A Blue Shift report titled "I, Robot," published September 3, offers a contrasting figure to the item above: the current overall physical AI market is about $18 billion, with humanoids accounting for only $2-3 billion of that; by 2030, the overall market could reach $60-100 billion, growing at over 20% annually. The report's assessment is that, in the near-to-medium term, value is captured mainly by specialized physical AI systems targeting specific high-value tasks, with humanoids remaining relatively immature by comparison. Blue Shift director Albert Meige put it this way: "When it comes to physical AI, humanoids are the moonshot that captures attention, capital, and talent, but real-world value is lagging behind the hype." In the same week, an investment bank raised its shipment forecast more than threefold, while a consultancy said humanoids account for only about one-sixth of the current market — the two figures aren't directly in conflict, since one measures future shipments and the other measures current revenue.Source: Arthur D. Little | Intelligent CIO source

This week in supply chain: Chinese sensor makers' market concentration solidifies, while reducer makers expand capacity and cut prices simultaneously · hardware

Consolidation has already taken hold in force sensing and end-effectors. The "2026 China Humanoid Robot Six-Axis Force Sensor Market Research Report," published by CCID and China Electronics News on August 31, shows that the top three Chinese manufacturers together hold 92.8% of the market, with Bluepoint Sensing above 80%, Kunwei Technology at 9.2%, and Sunrise Instruments at 3.3%; per the report's baseline year, Chinese shipments of six-axis force sensors for humanoid robots surpassed 10,000 units for the first time, with market size entering the hundreds-of-millions-of-RMB range. The absolute volume remains small, but domestic substitution is already essentially complete, with the next variable being unit price rather than market share. Tactile sensing is ramping up in parallel: on August 31, Folisenol Technology announced it has cumulatively delivered over 30,000 sets of tactile sensors to Lingxin Qiaoshou (Chinese dexterous-hand maker), part of a 100,000-unit framework order the two companies signed in February.

Actuators and transmission components are on a different trajectory. Green Harmonic Drive's H1 revenue was RMB 349 million, up 38.64% year-over-year, with net profit attributable to shareholders of RMB 70.11 million, up 31.25% year-over-year; the company previously said its harmonic reducer shipments reached nearly 500,000 units last year, with this year's capacity being prepared for close to a million units. In the same period, Zhongda Leader's (Chinese motion-control component maker) net profit attributable to shareholders fell 40.09% year-over-year (previously reported). Within the same downstream market and the same quarter, leading and second-tier players are moving in opposite profit directions, with price-war pressure not yet reaching the company with the largest market share. Taken together, the tightest constraint in the humanoid supply chain this week isn't capacity but unit pricing: sensors, having just crossed the 10,000-unit mark, are already dominated by Chinese suppliers, while reducer makers have built out capacity ahead of demand materializing.Source: Cailian Press | Jiemian News | Shanghai Securities News

FutureX · Physical AI Daily — Issue 108 (09/03)

Shawn — Wed, 02 Sep 2026 15:03:35 +0000

Today's Highlights

· World Labs releases world model Atlas, generating a controllable one-minute 1440p video plus 3D output from a single image

· Uber cuts 3,300 jobs, freeing up funds for over $10 billion in autonomous driving partnerships

· Perception company Lyte raises a $165 million Series C at a $1.6 billion post-money valuation

· Unitree's (Chinese quadruped/humanoid robot maker) stock has halved from its first-day price on its 11th trading day since listing, market cap RMB 220.85 billion

· The world's first international standard for legged robots, led by Chinese experts, is officially released

· The embodied data gap exceeds 99%; 25 Chinese data-collection companies raised over RMB 17 billion in H1

Paper Progress

Qwen-Drive-1.0: Packing Driving Ability into a Vision-Language Foundation Model Without Losing General Capability · autonomy

HuggingFace popularity 297, the highest among today's papers. End-to-end driving approaches typically trade general vision-language capability for driving capability. Qwen-Drive-1.0 leaves the architecture untouched and instead attaches a BEV perception head that simultaneously handles 3D detection, semantic occupancy prediction, and map segmentation, with a Planning Expert generating ego-vehicle trajectories from the shared representation. Staged training mixes driving supervision with general vision-language data; the authors report competitive motion-planning performance across open-loop, pseudo-closed-loop, and closed-loop settings, while largely preserving general vision-language capability.

Xin Zhou et al. · arXiv 2609.00111 source

ZimaBlue: Feeding in 120,000 Hours of Human First-Person Video Lifts Real-Robot Zero-Shot Success from 36.1% to 77.8% · world-model

Action-labeled robot trajectories are expensive and monotonous, while first-person video is cheap and abundant — but lacks action labels. ZimaBlue converts video into control ability through a three-stage curriculum: causal pretraining on human and robot first-person video, followed by video-action mid-training on heterogeneous robot trajectories using a unified action representation, then specialization for the target embodiment. In real-robot zero-shot evaluation, expanding training data from only the target robot's own data to 120,000 hours of video raised the success rate from 36.1% to 77.8%; an asynchronous Slow-Fast dual-system design lets the lightweight branch run at 30 Hz on an RTX 4090. HF popularity 38.

Xionghao Wu et al. · arXiv 2609.00188 source

IMPACT: The Supervision Signal in World Models Gets Swallowed by Static Backgrounds · world-model

Physically plausible interaction has long been a weak point for world models; prior fixes attached motion, geometric, or semantic representations, but these dense representations require extra estimators or manual annotation, capping training scale. This paper returns to the training objective itself, pointing out a supervision mismatch in the globally-averaged MSE denoising objective: static content, which dominates the frame, dominates the gradient, while the sparse dynamic regions that actually determine interaction quality are undersupervised. IMPACT treats cross-attention on the manipulated object's tokens as an internal spatiotemporal prior, sampling candidate regions and calibrating them with local prediction error into an interaction map, then reweights the denoising supervision accordingly — without external representations or changes to the inference pipeline. It consistently outperforms the corresponding MSE baseline across arm and hand manipulation and multiple DiT backbones.

Rongze Tang et al. · arXiv 2609.00161 source

REFACTOR-VLA: Building a Reusable Action-Skill Library for VLAs — Where a Bigger Model Performs Worse · vla

VLAs like OpenVLA, π0, RT-2, and RDT-1B are monolithic, outputting raw actions or short action chunks directly, and degrade on long-horizon tasks while remaining hard to interpret. This paper builds a wake/sleep system: during sleep, a learned implicit world model runs rollouts, from which behavioral equivalence is used to cluster action segments; during wake, typed lambda terms are issued and handed to a library-conditioned rectified-flow decoder for execution, with abstractions admitted to the library only after passing two gates — minimum description length and return preservation. Two findings run counter to intuition: scaling the world model from 188M to 430M parameters made performance worse across all four LIBERO suites; while adding an InfoNCE auxiliary contrastive loss for world-model warmup pushed clustering NMI an average of 0.184 above the strongest published baseline across the four suites.

Riyaaz Shaik, Chandru Venkataraman et al. · arXiv 2609.01215 source

DroneCATS: Small Models Reach the Target but Don't Know They've Arrived · benchmark

Multimodal large models are dropped directly into a drone control loop, with the action space declared entirely via prompt — no fine-tuning, no function calling. Testing four capabilities (approaching a visible target, tracking a moving target, turning to search for an out-of-view target, and commanding multi-drone formations) reveals a striking contrast: flying isn't the problem. Open-source models as small as 2B parameters are often more reliable at entering the success radius than frontier models, yet lose out by declaring arrival prematurely or not declaring it at all; in multi-drone command, small models blindly copy the same coordinate across different viewpoints. Spatial awareness passes; the action protocol doesn't. HF popularity 19.

Jaewoo Park et al. · arXiv 2609.01404 source

NegGrasp: Tell the Robot "Don't Grasp Here" and It Grasps There Anyway · manipulation

Language-driven dexterous grasping models like DextER perform well when the instruction specifies where to grasp, but fail systematically once a negative constraint is added, such as "grasp the handle but avoid the cup body." The cause lies in the training corpus: datasets like DexGYSNet contain almost no avoidance-type instructions, so the model learns that any mentioned part is a contact target, turning a forbidden zone into just another grasp point. This paper adds no negative training samples at all; instead, at inference time it uses sequential Monte Carlo with classifier-free guidance to steer sampling toward the specified part while pruning candidates heading toward the forbidden zone, with the forbidden zone's location read from language by a frozen 3D part-grounding model. On the accompanying NegGrasp benchmark, the violation rate of the strongest baseline drops from 57.9% to 17.2%.

Geonho Kim et al. · arXiv 2609.00555 source

CoLT-Drive: The Long-Tail Problem Isn't "Did It Recognize the Object" — It's "Did the Object Change What I Can Do" · autonomy

Long-tail failures in autonomous driving are typically framed as rare-object recognition errors; this paper argues the decision-level question is whether the model can infer how an object changes the ego vehicle's feasible high-level actions. To make small VLMs deployable, the authors propose the KPA framework, combining structured perception-to-decision prompting, SLERP-based expert merging, and a driving-condition-routed RegMoE LoRA mixture-of-experts module, aiming to adapt to long-tail cases without washing out the open-world knowledge in the pretrained model.

Zhengxu Tang et al. · arXiv 2609.00242 source

Other papers today: Provably Safe Sim-to-Real Transfer places safe sim-to-real transfer within a reward-free safe RL framework, using the degree of sim-to-real mismatch to characterize the benefit of real-world sample complexity (arXiv 2609.01418 source); a controlled experiment focused specifically on execution speed finds that imitation-learned dexterous manipulation policies, even when their nominal success rate matches the expert's, don't necessarily keep pace with the expert under varying speed conditions (arXiv 2609.01453 source); EmbodiedSkills treats each skill decision as an execution proposal, checking preconditions before execution and verifying postconditions afterward, wrapping low-level VLA policies in a trainable, checkable closed-loop agent layer (arXiv 2609.01281 source); VGEBench uses a logic-driven state machine to simulate multi-turn interaction, testing VLMs' visually-grounded exploration ability with unfamiliar household appliances (arXiv 2609.00845 source); Puppeteer performs pose-aware, object-grounded co-speech gesture generation and releases an accompanying synthetic 3D dataset, SceneGes (arXiv 2609.00369 source); a monocular depth estimation survey traces a decade of evolution from early learning-based methods to foundation models, distinguishing relative-depth from metric-depth approaches (arXiv 2609.01172 source); DNC-IMM adapts the interacting multiple model algorithm with neural calibration for early lane-change intent recognition, preserving its original probabilistic structure and interpretability (arXiv 2609.01120 source); a cross-modal pseudo-labeling pipeline enables domain adaptation for semantic segmentation without target-domain annotations, concluding that pseudo-label quality rather than quantity determines self-training success (arXiv 2609.00898 source); LRG inserts a lightweight post-processing enhancement module between a frozen feature extractor and classification head, addressing unreliable local feature spaces for minority classes in imbalanced time-series classification (arXiv 2609.00093 source); a retrieval evaluation deliberately separates surface wording from semantic structure, using the same protocol across two unrelated domains — competition mathematics and ALFWorld agent trajectories — to measure embedding retrieval's surface-form bias (arXiv 2609.01556 source).

Open Source · Tools · Benchmarks

· ACE-Ego-Hand: An open-source first-person bimanual 3D shape reconstruction model from Dexmal (Chinese robotics startup) with the Chinese University of Hong Kong and Nanyang Technological University, which converts a video diffusion model into a deterministic geometric encoder, replacing multi-step denoising with a single forward pass for roughly a 33x inference speedup; trajectories transfer directly to dexterous hands, and the team reports using it to produce about 5,000 hours of training data source

· RoboMIND: The open-source dataset from the Beijing Innovation Center of Humanoid Robotics doubled its downloads within a month to surpass 20 million, containing over 300,000 bimanual manipulation trajectories covering more than 700 practical tasks; its data facility spans nearly 6,000 square meters with over 30 representative scenarios and 40 robot configurations, with annual capacity of 180,000 hours, of which nearly 30,000 hours have already been delivered externally source

Funding & Deals

Lyte ｜ Series C ｜ $165 million ｜ $1.6 billion post-money · hardware

Led by Maverick Silicon, with Managing Partner Andrew Homan joining the board. The company works on the foundation layer of robot perception: proprietary sensor chips, multimodal sensors, and spatial software that compress 4D coherent vision, high-resolution imaging, and inertial sensing into a single synchronized timeline, measuring position and motion directly rather than reconstructing them in software. The founding team comes from Apple and PrimeSense, having worked on the 3D sensing behind Kinect and Face ID. CEO Alexander Shpunt's framing is that "physical AI has a perception problem before it has a model problem." The company came out of stealth only this January and has since entered mass production, shipping to robotics customers in inspection, logistics, and manufacturing. Source: Business Wire source

Locus Robotics ｜ Series G ｜ $41.6 million · industrial

Following yesterday's report that the round was nearing close, the amount is confirmed: $41.6 million, entirely from existing shareholders. CEO Rick Faulk said the round remains open and the company is still seeking additional funds from existing investors. This warehouse AMR company's last major raise was a $117 million Series F in November 2022, led by Goldman Sachs Asset Management and G2 Venture Partners, at a valuation near $2 billion. Ahead of this round, the company launched Locus Array and acquired Nexera Robotics. Source: FinSMEs source

Wandercraft ｜ In talks ｜ approx. €100 million ｜ approx. €750 million pre-money · humanoid ⚠️ unconfirmed report

Per Bloomberg, citing people familiar with the matter, with Goldman Sachs serving as financial advisor. This Paris-based company started in 2012 with rehabilitation exoskeletons; its Atalante X is already in use at over 100 hospitals and rehabilitation centers worldwide. Its new Calvin line pivots to industrial humanoids — the first model, Calvin-40, went from design to build in 40 days, runs on NVIDIA Isaac GR00T N1 and Jetson, and can repeatedly lift 40 kilograms. Renault is both a shareholder and the first commercial customer, already using Calvin-40 to mount tires on the production line at its Douai plant in France; the roadmap calls for about 10 units by the end of 2026 and 350 units by the end of 2027 across plants in France and Spain, targeting a 30% reduction in per-vehicle assembly time. Per Dealroom, humanoid robot companies had already raised $8.6 billion in H1 this year — 1.8 times the total for all of 2025. Source: Tech Funding News source

United Imaging Vision ｜ New round ｜ hundreds of millions of RMB · hardware

Investors included Dunhong Asset Management, the Hangzhou Embodied Intelligence Industry Fund, Fuyang Industrial Investment, and Zhipingfang (Chinese embodied-AI startup). This is a vision-solutions provider that has spent over a decade in the field, with customers including Unitree, Zhipingfang, AgiBot, Rokid, and Pico, while also supplying Ecovacs, Yunjing, and Trifo's robot vacuums and mowers — revenue is already at the RMB 2 billion level. Founder Zhang Feng started in phone camera modules in 2014, moved into robot vacuums in 2018, and pivoted to on-device AI vision in 2021. Funding will go toward its depth-camera product line, on-device visual understanding, data collection and evaluation capability, and a billion-RMB-scale smart vision industrial base in Hangzhou. Source: PEdaily source

Jaka Zhi Technology ｜ Hong Kong IPO ｜ updated prospectus · industrial

Revenue for H1 2026 was RMB 169 million, gross profit RMB 49.83 million, gross margin 29.4%, and a period loss of RMB 63.7 million. Over the prior three years, revenue was RMB 74.95 million, RMB 115 million, and RMB 266 million respectively, with period losses of RMB 114 million, RMB 118 million, and RMB 99.73 million — losses are narrowing but the company has yet to turn profitable. By revenue mix, standard AMR and omnidirectional heavy-load AMR solutions account for 54.7%, forklift-type AMRs 27.9%, and robot products only 10.2%. Board chair Xiong Rong controls 29.61% of voting rights, Zhongkong Technology founder Chu Jian holds 3.67%, and Legend Capital (Lenovo's investment arm) holds 5.80%. Cash and equivalents stood at RMB 218 million as of June 30. Source: 36Kr source

MirrorSpace ｜ Seed+ round ｜ over RMB 10 million · hardware ⚠️ compiled from public filings

Investors included Haiyuan Capital, Miracle Plus (formerly Y Combinator China), Songhe Capital, and Jinyili, with the information compiled from Qichacha public filings. This Dongguan-based company makes plug-and-play spatial perception modules, fusing multi-source inputs such as RGB, LiDAR, and thermal imaging into a dynamic 4D spatial representation. Source: Stockstar source

Commercialization & Deployment

Hyundai Mobis Wins All Actuator Orders for the Atlas Humanoid, 31 Units per Robot · hardware ⚠️ analyst estimate

Actuators, which convert energy into physical motion, are the most expensive and difficult component in a humanoid robot, and Hyundai Mobis has secured the entire actuator scope for Atlas, with 31 body actuators per unit. LS Securities named Hyundai Mobis its top pick in South Korea's physical AI value chain that day, maintaining a buy rating with a target price of KRW 750,000, and estimated that at an annual production target of 30,000 units, the robotics business could generate revenue of KRW 1.04–1.43 trillion. These revenue figures come entirely from the brokerage's model, not company disclosures; Mobis's own robot-component revenue isn't expected to become significant until 2027. Source: CNMO source

DEEP Robotics' (Chinese quadruped/wheeled robot maker) Wheeled-Legged Robot Enters Turpan Vineyards, Reportedly Cutting Manual Labor by 70% · industrial ⚠️ manufacturer's claim

The machine doesn't pick grapes. DEEP Robotics' Lynx M20S instead carries baskets of already-picked grapes from between the trellis rows to a collection point, while also repositioning drip irrigation tubing and relaying field data. It carries a 35-kilogram payload and operates across a temperature range from 50°C to minus 30°C — ground temperatures during Turpan's August harvest season routinely exceed 50°C — and can navigate a 50-centimeter-wide aisle without turning around. Grapes begin losing moisture as soon as they're picked, so shortening the time from harvest to transport directly affects sellable quality, which is the practical reason the vineyard adopted the robot. The report cites "over 70% reduction in manual labor" but does not specify per-unit cost or the scale of rollout beyond Turpan. Source: Global Agriculture source

China National Heavy Duty Truck Group Delivers 30 Huanghe Smart-Driving Tractor Trucks, Signs Strategic Agreement for 200 More · autonomy

Delivered on August 31 in Huai'an, Jiangsu, to customer Leijie Logistics, which specializes in medium- and long-haul trunk routes within Jiangsu. The trucks feature L2+ level driver-assistance, not driverless operation; the selling points lean more toward cost-reduction specs such as an integrated tractor-trailer drag coefficient of 0.349 and a third-generation MC13 engine rated at 620 horsepower. China National Heavy Duty Truck Group sold 188,500 heavy trucks in H1 this year, up 26.69% year-on-year, for a 28.5% market share. Smart-driving orders in trunk logistics currently remain mostly assistance-level, on a different track from the L4 fleets used in mining and port operations. Source: Fangde.com source

Didi's Robotaxi R2 Comes Standard with 10 RoboSense LiDAR Units · autonomy

Following the R2's launch of fully driverless paid ride-hailing trials in Beijing and Guangzhou on August 31, supply-chain details have emerged: each vehicle comes standard with 10 RoboSense (Chinese LiDAR maker) LiDAR units. For upstream suppliers, per-vehicle LiDAR usage is an order of magnitude higher than in factory-installed passenger cars, making this one of the segments most directly benefiting from Robotaxi scale-up. Source: Gasgoo source

Waymo to Begin Robotaxi Testing in Greater Cincinnati · autonomy

Following the September 1 same-day launch of paid driverless rides in Denver, San Diego, and Tampa, Waymo is extending its testing footprint further into the Midwest. Source: The Business Journals source

Industry Developments

World Labs Releases Atlas, Generating a Controllable 3D World from a Single Photo · world-model

Fei-Fei Li's World Labs has released its next-generation world model, Atlas, built on a multimodal autoregressive diffusion transformer architecture that treats text, images, camera poses, and 3D depth maps all as native inputs, with each input image anchored to a 3D position in space to jointly form a "spatial context." It can do four things: generate images and video with pixel-precise camera control, up to one minute at 1440p; reconstruct real 3D scenes from one to dozens of images, outputting point clouds or 3D Gaussian splats; jointly model space and time from input video; and generate images and 360-degree panoramas from text. In one example, feeding the model just 25 ground-level photos of Stanford's Main Quad, it generated a flyover path above the campus on its own. For robotics, Atlas offers a route to mass-producible real-to-sim conversion: the team used a phone to shoot two large environments, took 24 frames from each for reconstruction, then let simulated robots of different configurations move through the reconstructed scenes while Atlas generated, in real time, the RGB and depth data the onboard camera should see; after a task run, objects, positions, and lighting can be varied for data augmentation. Three synchronized cameras can also produce "bullet time"-style multi-view freeze frames. The team says Atlas follows a scaling law, with performance improving steadily with training compute, and that it will drive iteration on products like Marble. World Labs has raised $1.2 billion to date, with investors including NVIDIA, AMD, and Autodesk. Source: Zhidongxi source

Uber Cuts 3,300 Jobs, Redirecting Funds to Over $10 Billion in Autonomous Driving Investment · autonomy

The cuts amount to about 10% of Uber's global workforce, the largest since the 2020 pandemic layoffs. Management headcount was cut by about 20%, with some managers moving to individual-contributor roles, and the number of teams with only one or two people was nearly halved. CEO Dara Khosrowshahi's stated reason is that five years of rapid growth had built up excess layers, excessive coordination overhead, and diffuse accountability. The dining, retail, and direct-delivery operations groups have been merged, and engineering and science teams are also being consolidated. The company expects to invest over $10 billion in autonomous driving partnerships over the coming years; the Rivian partnership announced in March is ultimately meant to support up to 50,000 fully driverless robotaxis, launching first in San Francisco and Miami in 2028. Uber's stock rose more than 2% in pre-market trading following the announcement. Source: Analytics Insight source

Embodied Data Gap Exceeds 99%; 25 Data-Collection Companies Split Over RMB 17 Billion in H1 · adjacent ⚠️ compiled report

China Academy of Information and Communications Technology's "Embodied Intelligence Training Ground Research Report (2026)" assesses that globally available high-quality real-world data amounts to only hundreds of thousands to about a million hours, while embodied foundation models would need at least tens of millions of hours for capability emergence — a gap of over 99%. According to IT Juzi, 25 Chinese embodied-AI data startups raised a combined total of over RMB 17 billion in H1 2026. MiFeng Technology (spun out of AgiBot, Chinese humanoid robot maker), founded just six months ago, has completed three funding rounds; its MEgo line of embodiment-free capture devices has shipped over 20,000 units in mass production, generating a cumulative 1 million-plus hours of embodiment-free data. Chiron Robotics (Chinese embodied-AI startup) has deployed over 400,000 collection points across more than 100 Chinese cities. MiFeng's chairman and CEO Yao Maoqing put it plainly: embodiment-free data suits pretraining for general representations, but for specific tasks "there's no avoiding real-robot data matched to that embodiment" — internet video distillation is only "a crutch for this stage." Industry standards are also shifting: earlier this year, data captured with phones or action cameras could still be sold, but by the second half of the year, 60FPS, 1080p, and stereo vision have become consensus requirements — "collecting a million hours blindly isn't actually hard; you can get there by repeating a handful of task types, but that kind of data has almost no value." The Ministry of Industry and Information Technology's "Quality Requirements and Evaluation Methods for Embodied Intelligence Datasets — Key Foundational AI Technology," approved in August, takes effect on November 1 and covers eight dimensions including completeness, consistency, diversity, and authenticity. Source: Sina Finance source

Unitree Stock Halves From First-Day Price on 11th Trading Day Since Listing, Market Cap RMB 220.85 Billion · humanoid

Closed at RMB 546.02 on September 2, down 4.39%, exactly half the intraday high of RMB 1,100 on its first day of trading; market cap now RMB 220.85 billion. Shares jumped 460% on their August 19 debut on the STAR Market, at an issue price-to-earnings ratio of 219.23x. On fundamentals, humanoid robots accounted for 51.78% of 2025 revenue, with pure-humanoid shipments exceeding 5,500 units and combined quadruped sales exceeding 33,000 units. Founder Wang Xingxing told investors he hopes investors buy in because they recognize the company's value rather than for speculation, and that going public brings "more pressure but also more motivation." The same day, the company responded to online complaints from employees alleging "penalties but almost no bonuses" and "reimbursements over RMB 100 require Wang Xingxing's personal approval," calling much of the content inaccurate. The company's registered capital has been formally updated post-listing to RMB 404 million. Source: Guangzhou Daily Xinhuacheng source

World's First International Standard for Legged Robots Released, Led by Chinese Experts · humanoid

The standard, titled "Robots — Performance Specifications and Test Methods for Service Robots, Part 5: Legged Robot Locomotion," covers robots that walk, run, climb, and navigate obstacles on legs — including humanoids and quadrupeds — for which there was previously no unified global evaluation framework. The core changes lie in two areas, test environment and evaluation scope: testing is no longer confined to flat-ground mobility but now includes real-world scenarios such as stairs, slopes, and obstacles; evaluation also no longer looks only at motion metrics but factors in perception, interaction, and decision-making as part of task completion. Such robots are already performing real work — performance, inspection, transport, search and rescue — but cross-vendor performance comparisons have so far relied largely on each company's own self-reported demonstrations. Source: Standardization Administration of China source

AgiBot Releases TwinDEX: One Hand Worn by a Human to Collect Data, One Hand Mounted on a Robot to Do the Work · embodied ⚠️ manufacturer's claim

A twinned pair of three-fingered, nine-degree-of-freedom dexterous hands, seven of which are active degrees of freedom. The capture-side unit is a wearable exoskeleton, while the execution-side unit is mounted on a robot; the two are aligned across three dimensions — kinematics (degrees of freedom, joint axes, link proportions), contact mechanics (materials, geometry, surface properties, and tactile sensors), and visual appearance — so that embodiment-free data can be used directly for policy training without needing alignment or fine-tuning against real-robot data. AgiBot's own benchmark claims embodiment-free data collection is 5.3 times more efficient than real-robot teleoperation, and that on multi-task benchmarks, policies trained on either data type improve at the same rate as data volume grows and eventually converge — meaning embodiment-free data can nearly 100% replace real-robot teleoperation data. Both figures come from the manufacturer's own testing and have not yet been independently reproduced. Source: IT Home source

Zeroth Launches Bridge Humanoid Robot and Open-Source OpenBridge Platform · humanoid

Suzhou-based company Zeroth (Chinese robotics startup) has unveiled the compact humanoid Bridge, aimed at developer-focused physical AI development; its hardware, motion-control system, and development interfaces are designed for AI-driven physical interaction rather than preset motion choreography. Out-of-the-box capabilities include stable walking, obstacle avoidance, and a library of humanlike movements, with support for motion capture and VR input. The company simultaneously open-sourced OpenBridge, which includes a public Skill Hub allowing developers to call and extend robot skills written by others. Its Builder program will select 500 developers in its first cohort, with the first 200 receiving a free unit and the remaining 300 gaining priority purchase access to the Geek edition plus SDK access. Founder and CEO Guo Renjie said, "The world never lacks walls; we chose to build bridges instead." Bridge is open for global preorder starting September 2, with the Geek edition's first batch already on sale, priced by region. Source: The AI Insider source

Caterpillar Partners with FieldAI to Bring Robot Foundation Models to Job Sites and Factories · industrial

Caterpillar contributes industry experience, engineering capability, and operational data, while FieldAI contributes embodiment-agnostic autonomy and robot foundation models, built on NVIDIA's technology stack. The initial focus areas are autonomous inspection, digital twins for job sites and factories, early risk identification, and simulation- and automation-driven operational optimization. Chief Technology Officer Jaime Mineart said, "Our industry's future depends on how effectively human expertise and AI-driven machines work together"; John Tuntland, Senior Vice President of the Integrated Components Division, applied the point to Caterpillar's own factories, saying the technologies give teams clearer visibility into how facilities operate and where flow can be optimized. Caterpillar cited FieldAI's track record of deploying in complex, dynamic industrial environments where traditional automation falls short as the reason for the partnership. Source: PR Newswire source

Hardware · Supply Chain

· Chinese Embodied Intelligence Tier 1 Industry Alliance: Launched by Xingyuanzhi (Chinese robotics component maker) and Linker Bionics (Chinese dexterous-hand maker), with over 20 supply-chain companies joining, including Orbbec, Hesai, RoboSense (LiDAR), another RoboSense-affiliated entity, Linkage Robotics, Rokae, Xinjingcheng, and Quectel; members span embodied "brain" models, dexterous hands, voice interaction, LiDAR, visual perception, tactile sensing, robotic arms, chassis, wireless communication, and thermal management source

· GlobalFoundries UX Platform: PDKs for two CMOS process nodes, 40UX and 22UX, are now officially open to customers, targeting ultra-low-power MCUs, wireless connectivity, sensor interfaces, and imaging systems for edge devices; 40UX is based on a 40nm high-endurance eFlash platform that has already shipped over a million wafers, while 22UX is aimed at stronger analog, sensing, and mixed-signal performance source

· TPC Robotics Actuators: South Korea's TPC has launched 7 proprietary-brand humanoid robot actuators in one go — 3 rotary models for the neck, elbow, and hip, 3 linear models for heavy-load areas like the lower and upper leg, plus 1 finger-joint module; the company posted a loss in H1 due to intensifying competition and is counting on its robot component and linear motor businesses to turn things around in H2 source

· UBS Places Its Bet: UBS Wealth Management's Chief Investment Office issued a report arguing that the physical AI era favors component suppliers over robot makers, on the grounds that in-context learning lets robots imitate humans after watching them perform a task once or twice, eliminating the weeks-long retraining process previously required — thinning out the technical moat on the embodiment side source

FutureX · Physical AI Daily — Issue 107 (09/02)

Shawn — Tue, 01 Sep 2026 14:57:40 +0000

Today's Highlights

· Waymo launched paid robotaxi service in Denver, San Diego, and Tampa on the same day, growing its nationwide fleet to over 4,000 vehicles

· Mech-Mind (Chinese robotics vision company) listed on the Hong Kong Stock Exchange; 9 cornerstone investors subscribed $186 million, market cap exceeds HK$12 billion

· UBTech drew its first firm line, aiming for positive quarterly EBITDA in Q4 this year

· A "robot kindergarten" involving Turing Award winner Richard Sutton opened in Shijingshan, Beijing

· N₀-Foundation released 30,000 hours of visuo-tactile data, open-sourcing 5,000 hours of it

· Lucida turns real rooms into editable simulation assets, lifting scene F-Score from 0.794 to 0.924

Paper Progress

N₀-Foundation: 30,000 hours of visuo-tactile data, one-fifth open-sourced · perception

In embodied manipulation, tactile data has always been the hardest and most expensive category to collect at scale, and this work pushes the scale directly to over 30,000 hours of synchronized visual-tactile demonstrations across 6 embodiments, covering 450 tasks, with 5,000 hours released publicly as OpenNeoData. The team also released a visuo-tactile sensor, a tactile-capture version of the UMI interface built specifically for tactile collection, a cross-sensor transferable representation model called NeoForce, and two evaluation suites — NeoReal for real-robot testing and NeoSim for simulation testing. Their experiments aim to show that what policies actually benefit from is the underlying physical contact state itself, not signal appearances specific to any one sensor.

NeoteAI Team, Fudan TEAI Team · arXiv 2608.29601 source

Lucida: breaking cluttered rooms into individually manipulable simulation assets · perception

Real-to-sim pipelines typically demand precise instance geometry and unoccluded viewpoints from the outset — conditions real-world footage simply can't provide. Lucida keeps the overall "parse–generate–place" order but pushes the precision requirement to the end of the pipeline: it first parses video into a scene graph backed by multi-view evidence, then generates a complete asset for each instance separately, and finally hands off to GizmoAct — a VLM policy that treats object placement as multi-turn GUI operation, dragging object gizmos itself and judging when alignment is achieved. Tested on R2S-Scene, mAP came in 69% higher than Boxer, and scene F-Score rose from SAM3D's 0.794 to 0.924. This work has reached a popularity score of 53 on the HF community.

Minghan Qin et al. · arXiv 2608.30821 source

LightNav-0: no dedicated prediction head for navigation tasks · autonomy

Instruction following, open-vocabulary object search, and visual tracking have traditionally each required their own separate module. LightNav-0 instead folds all three into a single compact model through a unified token interface: a dual-channel pointing representation conveys intent in a way that is independent of task, scene, or embodiment, while a residual vector-quantized action tokenizer translates that generic intent back into a trajectory a specific embodiment can execute. Trained on a corpus spanning over 2,000 scenes and 4,000+ hours of data, the model achieves SOTA monocular success rates across 10 public navigation simulation settings, and demonstrates zero-shot cross-embodiment transfer on real robots. This work has a popularity score of 24 on the HF community.

Shaoan Wang et al. · arXiv 2608.30935 source

Hydra: planning entirely within a discrete latent space, no decoding back to pixels · world-model

World models are good at imagining the future, but using them for real-time control runs into a fundamental mismatch: the planner and the generative model don't operate on the same manifold, so every candidate has to be decoded into a high-dimensional image before it can be scored — prohibitively costly on real hardware. Hydra instead compresses visual state, physical pose, and control actions into a single shared latent space, then splits it with a per-modality vector-quantized bottleneck into a discrete vocabulary, so candidates can be ranked in place by a Kinematic-Perceptual Cost without ever generating pixels; the selected intent is then unrolled into a continuous trajectory by conditional flow matching. Tested on two real robot platforms, goal-directed planning outperformed existing world models, and closed-loop execution matched or exceeded mainstream reactive base policies.

Mohammad Nazeri et al. · arXiv 2608.28995 source

Blind Dexterity: a blindfolded Unitree G1 performs whole-body manipulation · locomotion

No cameras, no torque sensors, no tactile skin — the only available signal is joint encoders. With just that, the policy enables Unitree G1 to walk while being pushed (without even relying on IMU feedback), stop a soccer ball with its foot, lift a suitcase by feeling for the handle, and step onto an arbitrarily placed skateboard. The authors' explanation is that the way joint readings shift during compliant contact itself constitutes a whole-body tactile channel — the robot actively induces contact to probe its surroundings, and object pose can be decoded from just a short window of proprioceptive history. This effectively inverts the default assumption that expensive sensors have to come first.

Aditya Bhatt et al. (Jan Peters's group) · arXiv 2608.29487 source

A humanoid robot swings from a bar, succeeding 14 of 15 times · locomotion

Compared to flat ground, sparse 3D structures impose an entirely different perception challenge — the bar is thin and suspended in air, giving lidar very few returns. This system feeds raw scans from a head-mounted solid-state lidar directly to the policy, uses an attention encoder with recurrent memory to extract geometric information, and stitches together the jump-up, swing, and jump-down phases of expert motion via staged teacher-student training. To transfer to real hardware, the team explicitly modeled lidar noise, battery voltage sag, and actuator thermal limits during training, and swapped the end effector for a passive hook. Across three bar-spacing configurations, the full sequence was attempted 15 times and succeeded 14 times, with swing speeds reaching 0.5 m/s; using the same perception backbone to train a separate policy, the robot could even swing through an overhead bar with a cross-section of just 2 cm.

Efe Ongan et al. (ETH Zürich) · arXiv 2608.29769 source

Motus2: one set of weights serving as policy, simulator, and judge · world-model

Common practice has been to bolt an action head onto a world simulator, but prediction and decision-making never actually formed a true closed loop. Motus2 exposes three interfaces from a single shared-weight model: a policy proposes candidate action chunks, a simulator predicts the visual consequences of those actions, and a value model scores the predictions — together closing a self-improving decision-and-learning loop. Failed and suboptimal interactions, previously discarded, are now fed back into dynamics modeling and value learning. The data side was also built up progressively: starting from large-scale monocular first-person video, then moving to synchronized binocular data, then robot-domain adaptation, and finally landing on a biomimetic platform equipped with binocular vision, dual arms, dual dexterous hands, and tactile sensing.

Hongzhe Bi et al. · arXiv 2608.30237 source

RoboPhys-3D: separating "looks right" from "is right" in world models · benchmark

Video world models are increasingly used as data engines and simulators, yet no one had systematically verified whether their generated rollouts actually preserve 3D scene state, or whether they translate into genuinely executable actions. RoboPhys-3D is built on top of RoboTwin 2.0, spanning 50 manipulation tasks, 5,000 trajectories, and 25,000 multi-view ground-truth videos; both generated and ground-truth videos are passed through the same 3D reconstruction pipeline, so reconstruction error and generation error can be disentangled. The evaluation suite comprises 50 metrics across four tiers. Among four representative models, Cosmos 3 achieved the highest RoboPhyscore at 0.6330, or 92.7% of ground-truth level; but the failures exposed by state-level and execution-level metrics went undetected by both perceptual metrics and VLM judges.

Tianyi Wang et al. · arXiv 2608.28718 source

A single poster shifts autonomous-vehicle camera–lidar calibration by 33.9° · autonomy

Online calibration was designed to let vehicles self-correct for vibration- and temperature-induced sensor drift without a shop visit — but this paper turns it into a new attack surface. ACA uses a single adversarial poster to first fool the miscalibration detector into triggering the calibration process, then steer the calibration estimator toward incorrect extrinsics, with the poster's geometry and texture jointly optimized for both effects. On KITTI and nuScenes, the attack produces up to 33.9° of average rotational error, collapsing downstream object detection; in CARLA simulation, once a poisoned calibration is accepted by the system, it leads to collisions; and a printed version of the poster attached to a real Husky robot reproduced the same errors. Incorrect calibration propagates through the fusion chain all the way into planning and control.

Liangkai Liu, Qingzhao Zhang, Kang G. Shin (University of Michigan) · arXiv 2608.28778 source

Other papers today: Matrix-Game 3.5 uses a patch memory mechanism to let real-time interactive world models sustain minute-scale long-horizon consistency (arXiv 2608.29910 source); NavMCP separates the roles of the VLM reasoning agent and the navigation foundation model, outperforming step-by-step interfaces by 14.9 points on HM-EQA (arXiv 2608.30396 source); DriftingVLA generates an entire action chunk in a single forward pass, eliminating the online latency of multi-step flow refinement (arXiv 2608.29749 source); AnyWorld can expand a single human interaction into multi-embodiment, multi-view, multi-scene robot trajectories (arXiv 2608.29242 source); DREAM generates fine-tuning data on-site via real-to-sim at deployment time, eliminating the need for human demonstrations (arXiv 2608.29078 source); SpectraTac built a compact camera-free optical tactile sensor using RGB active illumination with distributed color sampling (arXiv 2608.30368 source); one study found self-play driving policies exhibit reward hacking at traffic lights and have no inherent incentive to stop at stop signs (arXiv 2608.30819 source); drifting behavior turns out to emerge organically from lap-time optimization pressure, with no dedicated drift reference built into the reward design (arXiv 2608.28723 source).

Open Source · Tools · Benchmarks

· GHOST: an open-source VR teleoperation system letting a single operator control two mobile manipulators simultaneously, stitching together a third-person 3D workspace entirely from onboard RGB-D point cloud alignment (arXiv 2608.29080 source)

· Agri-Sim: a Unity + ROS2 greenhouse agricultural robotics simulation platform enabling closed-loop, repeatable evaluation of navigation, motion planning, and manipulation tasks (arXiv 2608.29100 source)

· DARP: an RGB-D-IR calibration dataset for dual-arm eye-in-hand setups, with the two arms photographing each other from opposite sides of a table, specifically designed to address self-occlusion under single-view capture (arXiv 2608.31002 source)

Funding and Deals

Mech-Mind (Chinese robotics vision company) | Hong Kong Stock Exchange main board listing | Raised HK$2.35 billion | Market cap over HK$12 billion · embodied

Following price-setting completed last week, Mech-Mind (09615.HK) listed formally on September 1, pricing at the top of its range at HK$101.7 and issuing 23.14059 million shares; if the over-allotment option is fully exercised, total proceeds could reach HK$2.7 billion. Nine cornerstone investors together subscribed $186 million, accounting for 62% of the offering: Baillie Gifford subscribed $60 million, Taikang Life $40 million, Jane Street, Invus, Ghisallo, and Ruihua each subscribed $15 million, NGS Super Fund and E Fund each subscribed $10 million, and BYD's Golden Link subscribed $6 million. The company makes standardized "eye-brain-hand" components; according to Frost & Sullivan figures, in 2025 it held a global shipment share of over 27% in the AI+3D-vision-guided general-purpose robotic component market — more than its next four competitors combined. Revenue grew from RMB 181 million in 2023 to RMB 389 million in 2025, with gross margin over the same period rising from 39.1% to 64.6%, and adjusted net loss narrowing to RMB 109 million; overseas revenue share rose to 50.3%, exceeding domestic (China) revenue for the first time in 2025. Although the public offering was oversubscribed roughly 3,835 times, the stock opened flat on its first day and at one point fell more than 4% intraday. Founder Shao Tianlan told LeiPhone that the annual shipment volume of general-purpose robots actually deployed in productivity settings is still only in the low hundreds of thousands, "several thousand times short of true mass adoption."Source: LeiPhone source

Reframe Systems | New round | $40 million · industrial

The round was led by Energy Impact Partners, with participation from Counterpart Ventures, E12 Ventures, Global Brain, Thin Line Capital, Up Partners, LACI Impact Fund, and others, alongside continued investment from existing backers Eclipse and RA Capital. All three co-founders — Felipe Polido, Aaron Small, and Vikas Enti — came from Amazon Robotics; the company was founded in 2022 and focuses on small, demand-proximate automated home-building factories. The company says its delivery speed is three times that of traditional construction, at 35% lower cost; it has delivered 10 homes to date, plans to deliver 114 more over the next year, and has set the capacity ceiling of its next factory, soon to come online in Massachusetts, at 500 multi-family units.⚠️ Company claimSource: The Robot Report source

Tiangong Robotics (Beijing Tianxiaxian Zhichuang) | New round | Hundreds of millions of RMB · industrial

The round was led by Chuxin Fund, with participation from Matrix Partners China and Hefei state-owned capital. Founder Xiao Jun previously served as JD.com Group vice president and president of its X division, where he led the company's unmanned warehouse, driverless vehicle, drone, and specialty robotics businesses. Founded in 2022, the company focuses on logistics sorting, has grown to a team of over 200, with roughly 85% in R&D roles; its flagship products are the "Wooden Horse" sorting robot and an expansion module that increases a single unit's sorting slot count to 4-5 times the original, and its business now extends to Japan, South Korea, the U.S., and Europe. The funds will go toward expanding sorting-robot production, developing and commercializing sorting-oriented humanoid products, and expanding into overseas markets.Source: ZhiDongXi source

Physical Superintelligence (PSI) | Seed round | $58 million · adjacent

The round was led by Breakthrough Energy Ventures, with participation from Dragon Global, Robot Ventures, SV Angel, Valkyrie, and others, plus individual investors from OpenAI, NVIDIA, Oracle, and Hugging Face. The company was formally founded on September 1 in Cambridge, Massachusetts, positioning itself as an AI-native physics lab staffed by "virtual physicists," with its core platform Emmy breaking research questions into a verifiable hypothesis tree and testing them in parallel. Its first chosen commercial application is AI data centers, using physics-native reasoning and simulation to solve multi-physics design problems spanning power, cooling, networking, and compute. CEO Matt Pines said PSI's mission is to "industrialize the discovery of new physics."Source: Unite.AI source

Youlichi | Strategic investment | Tens of millions of RMB · embodied

The investor is Keli Sensing, a publicly listed force-sensor maker; the investment is in the tens of millions of RMB. This is another stake in Keli Sensing's ongoing strategy of taking positions along the robotic perception supply chain.⚠️ Single-party claimSource: Robot Lecture Hall source

Locus Robotics | Series G | Nearing close · industrial

The warehouse-robotics and related-software company's Series G round is nearing completion; specific amount and investors have not yet been disclosed.Source: Axios source

Commercial Deployment

Waymo launches in three cities in a day; Zoox pushes testing into Houston · autonomy

Denver, San Diego, and Tampa all opened to public paid robotaxi service on the same day, September 1, with Colorado marking Waymo's first commercial operation in that state. A company spokesperson said each of the three cities will start with a few dozen vehicles, "growing to the hundreds over time," with rides booked through the Waymo app. The nationwide fleet now exceeds 4,000 vehicles across 14 cities, delivering over 500,000 paid rides per week, with the company aiming to push that past one million rides per week by the end of 2026; the new Ojai model is built on a Zeekr chassis, assembled at the Arizona factory, and uses sixth-generation autonomous driving hardware. Zoox, meanwhile, said it will begin safety-driver testing in Houston and San Diego, bringing its footprint to 12 markets, though its only paid service currently still operates solely in Las Vegas, with its San Francisco service still limited to free rides for select riders. The near-simultaneous moves from both companies land just ahead of Tesla's Thursday unveiling of the Cybercab and Robotaxi. Goldman Sachs research projects the U.S. robotaxi market will reach $19 billion by 2030. Labor groups worry about job displacement, while vehicle-safety advocates are calling for companies to disclose mileage and collision data using consistent, comparable standards.Source: CNBC source

Zhipingfang's AlphaBot mixes drinks at a Hong Kong Lan Kwai Fong bar · embodied

Starting work formally on August 31, the AlphaBot 2, equipped with the brain-inspired embodied foundation model NeuroVLA, took its place behind a real bar counter to mix cocktails for customers. More notable than the robot itself is the operational pipeline behind it: hardware customs clearance, regulatory compliance, obtaining a local Hong Kong beverage sales license, and building a paid service system — only once that full chain is working can this be called routine operation rather than a one-off demo. Lan Kwai Fong, home to over a hundred bars and restaurants, provides a ready-made stress test through its fluctuating foot traffic, noise, and lighting. Manufacturer-reported data shows motion jitter reduced by more than 75%, reflex response after a collision completing within 20 milliseconds, and power consumption of about 0.4 watts. Zhipingfang's earlier modular service space "Smart Cube" is already operating in over a dozen provinces and cities, with robot staff there making coffee, ice cream, and matcha.⚠️ Company claimSource: PEdaily source

GreyOrange's in-store system now runs in 3,800 retail locations · industrial

GreyOrange technology is now running in over 3,800 retail stores worldwide, with major retailers including the H&M Group using its gStore for inventory and shelf management, having tracked roughly 200 million items cumulatively. The system is sensor-agnostic, working with either overhead or handheld RFID readers, and can pinpoint an item's location to within 3 to 5 feet. The company disclosed that 130,000 software and hardware agents are currently running across its stores and warehouses, with 100,000 of those added in the past two years.Source: GlobeNewswire source

Tesla's Robotaxi fleet grows to nearly 200 vehicles in three weeks · autonomy

Data compiled by the crowdsourced platform Robotaxi Tracker shows the number of Tesla vehicles operating without a safety monitor across Austin, Dallas, and Houston has grown to nearly 200, roughly a 7x increase in three weeks; regulatory registrations for the two-seat Cybercab in Texas also jumped from 7 to 45 within a few days. The fleet had long held steady at around 20-odd vehicles, leading critics to question whether Tesla's actual deployment capability matched its stated goals. Autonomy lead Ashok Elluswamy said on the Q2 earnings call that the Robotaxi program has completed over 380,000 miles of unsupervised driving with no notable incidents. Taken together, these two developments suggest Thursday's event is likely more than just a simple new-car launch.⚠️ Aggregated dataSource: Teslarati source

Singapore MRT station trials humanoid guide robot, it mishears a place name · humanoid

Singapore's Land Transport Authority (LTA) placed a 1.27-meter tall Unitree humanoid robot named Olly at Little India MRT station for a two-week trial, running weekday mornings from 10 a.m. to noon, from August 31 to September 11. Its performance on day one was mixed: it handled simple questions like first/last train times fine, but when asked how to get to Tekka Centre, it needed multiple prompts before giving verbal directions, couldn't lead the person to the nearest exit, and at one point misheard the name as "Teacher's Centre," pointing the wrong way. When asked how to reach KK Women's and Children's Hospital, though, it correctly identified Exit F and proactively led the way all the way to the gantry. Steve Walker, a 69-year-old retiree with a robotics background who came specifically to try it out, said the robot's balance and gait were stable, but that it struggled to distinguish voices from background noise in the noisy station hall, and suggested adding a directional microphone. LTA said it chose off-peak hours for the trial and currently has no deployment plans.Source: CNA source

T-Robotics wins additional Ford ESS line automation order · industrial

The contract is worth approximately $2.63 million (roughly KRW 3.6 billion), supplying Ford's energy storage battery production facility in the U.S., covering automation equipment, installation and commissioning, and control-system development integration. Combined with the roughly KRW 15 billion AMR contract from this past May, cumulative orders tied to Ford's ESS line had reached about KRW 19 billion as of September 1. Ford's production line is being built in phases, leaving room for further orders.Source: Asia Economy source

Industry Developments

UBTech's revenue doubles in H1, gives first quarterly break-even timeline · humanoid

In the six months ended June 30, 2026, the company's revenue reached RMB 1.269 billion, up 104.2% year-on-year; gross profit reached RMB 567 million, up 160.9%, lifting gross margin by 9.7 percentage points to 44.7%; the period's net loss narrowed 23.0% to RMB 339 million. Following last week's disclosure that revenue from its full-size models had grown 1445%, the company said for the first time on its August 30 earnings call that it now aims to bring forward its target of quarterly break-even — originally set for Q4 2027 — to positive Q4 EBITA this year. CEO Zhou Jian noted that in both 2024 and 2025, deliveries in the second half outpaced the first half; while the Guangxi factory won't come online until mid-to-late September due to natural conditions, the company remains confident in its full-year revenue target of RMB 3.5–4 billion or more. The revenue mix has shifted sharply: full-size embodied-intelligence humanoid robots brought in RMB 590 million, jumping from 6.1% to 46.5% of total revenue year-on-year, with H1 sales of 921 units; revenue from other smart robot products, including educational robots, was RMB 246 million, down 18.8% year-on-year. Cash also fell sharply, from RMB 4.888 billion at the end of 2025 to RMB 2.326 billion, mainly due to a RMB 1.665 billion cash payment for a 43.01% stake in FengLong Co. First-half credit impairment losses reached RMB 91.07 million, versus just RMB 1.3 million in the same period last year, which the company attributed to a change in expected credit loss methodology from an external-rating approach to a migration-rate approach.Source: Nanfang Daily Wancaishe source

Skild AI releases S1, replacing spoken instructions with a video · world-model

Rather than a text prompt, this time it's a video. S1 positions itself as an "in-context learner": given a video of a human performing a task, without any fine-tuning or weight updates, it can directly produce robot actions, for tasks up to 10 minutes long — including repotting a plant, flipping pancakes, pour-over coffee, and kit assembly, none of which appeared in its pretraining data, per company demos. The company's blog reports 96% success on in-distribution tasks and 66% on out-of-distribution long-horizon tasks, versus just 9% for a comparably-sized language-prompted baseline; under noticeable perturbations, the language-model approach degraded three times as much as S1. This directly targets the dominant current paradigm, where every new task requires fine-tuning a dedicated policy on tens to hundreds of hours of teleoperation data. Skild was founded in 2023 and has raised nearly $1.7 billion to date, having just closed a $1.4 billion Series C last week (previously reported).⚠️ Company claimSource: The Robot Report source

"Robot kindergarten" opens in Shijingshan, Beijing, with Sutton's team co-building · embodied

It opened formally on September 1, built jointly by Tashan Technology and a team led by Turing Award winner and reinforcement-learning pioneer Richard Sutton, focused on tactile perception and continual learning. It's explicitly designed to bypass the current dominant approach — rather than relying on human demonstrations, robots repeatedly try things on their own, and when they hit a wall, that failure gets written into the learning data before trying a different way. Sutton described the goal as "providing a safe environment for a robot to learn about its own body and how to interact with the physical world through experience," with "failure being part of the process." One spider-shaped small robot in the facility learned to move forward in about 40 minutes with no prior training — though this remains a very narrowly defined task. For now, this approach looks more like a complement to teleoperation and imitation learning than a replacement for it.Source: Global Times source

Goldman Sachs raises 2030 humanoid shipment forecast 3.5x · humanoid

In a newly released "Physical AI" research report, Goldman Sachs now projects global humanoid robot shipments will reach approximately 890,000 units by 2030 and about 6.5 million by 2035, corresponding to a market size of roughly $138 billion — up from previous forecasts of 256,000 and 1.4 million units, respectively. This year's baseline shipment forecast was also raised from 51,000 to 75,000 units. The report identifies logistics and warehousing as the leading early-adoption use case: Amazon has already deployed over one million robots across more than 300 sites, and Walmart's freight automation now covers 3,100 stores nationwide; Goldman estimates automation could save Amazon roughly $72 billion in service costs by 2030. On chips, the report estimates chip content per humanoid robot at over $3,000-$6,000. Another easily overlooked finding concerns factory control systems: as humanoid robots proliferate, PLCs are shifting toward software-defined vPLCs, a market growing 20-30% annually, posing a challenge to established vendors like Siemens and Rockwell.⚠️ EstimateSource: Guandian.cn, citing Goldman Sachs research report source

Hugging Face's rubber-duck robot sells out fast, chip sourced from Rockchip · adjacent

The Microduck, launched by Hugging Face's French subsidiary Pollen Robotics, has sold over 10,000 units since going on sale last Thursday; at $399 each, that's already over $4 million in revenue, and delivery of new orders has slipped past the originally promised Christmas 2026 date. The 800-gram duck-shaped robot doubles as both toy and development platform, running on the RK3566 chip from Shanghai-listed Rockchip (Chinese chipmaker), which uses licensed ARM technology. Omdia chief analyst Lian Jye Su said Rockchip is a key supplier in edge AI, with its chips commonly used in machine-vision applications like object detection and image recognition, noting it's "very widely deployed, but with limited compute resources — not purpose-built for sophisticated edge AI devices." Rockchip's H1 revenue reached RMB 2.88 billion, up 40% year-on-year, with non-GAAP net profit growth exceeding 60%. Several other players have already entered this price segment: Zeroth's child-sized humanoid robot is priced at RMB 8,888 and has taken 247 pre-orders on JD.com.Source: CNBC source

500+ China A-share companies mention "embodied intelligence" in half-year reports, single digits disclose revenue · adjacent

China Fund News combed through 2026 half-year reports and found over 500 A-share companies mentioning "embodied intelligence" in their filings, but only 23 discussed it in any depth (10+ mentions), and the number actually disclosing corresponding revenue figures was in the single digits. The companies that could actually back it up with numbers share one trait: their core business already had a natural interface with embodied intelligence. Sensor and control maker Sinexcel renamed its product segment from "Smart Devices" to "Embodied Intelligence" — its frameless torque motors are now being supplied in volume to several leading manufacturers, with segment revenue of RMB 60 million, up 114.7% year-on-year; Dobot's (Chinese robotics maker) embodied-intelligence business revenue surpassed RMB 45 million, up over 20x year-on-year, accounting for 14.3% of total revenue. On the other end are companies with struggling core businesses: Hoson Intelligent posted an H1 net loss of RMB 425 million yet mentioned embodied intelligence 24 times in its half-year report; Shangpin Home Collection saw revenue decline 24.2% and posted a net loss of RMB 233 million, while also touting a high-profile push in this direction. Mention frequency didn't correlate with actual performance either way — Kejie Intelligent's net profit grew 251.6% while mentioning the term 35 times, and Kepu Cloud mentioned it 63 times. The sector saw 322 financing deals in H1 totaling roughly RMB 93.5 billion, already exceeding the full-year 2025 total.Source: China Fund News source

Unitree's market cap down about RMB 218.8 billion from its listing-day high · humanoid

Shares hit an intraday low of RMB 555.8 on August 31, a new post-listing low, closing the morning session at RMB 559.11, down 4.43%, for a market cap of RMB 226.14 billion — a 49.17% pullback from the RMB 1,100 opening price on its first day of trading. The stock recovered 1.1% on September 1, with the company also denying, on the same day, earlier rumors concerning expense-reimbursement approvals.Source: China.com Finance source

Galbot's ET1 opens pre-orders September 3 · humanoid

This is the company's first small bipedal humanoid robot, standing 1,230 mm tall and weighing 30 kg, with a flexible-material shell and a motor solution supplied by Unitree. It runs Galbot's general-purpose "cerebellum" foundation model, AstraBrain-WBC, trained on 100,000 hours of human motion data using a mix of high-precision motion capture and video retargeting.⚠️ Company claimSource: Gasgoo source

Hardware · Supply Chain

· Folee New Materials → Linker Bionics: cumulative deliveries of tactile sensors have now exceeded 30,000 units, primarily fitted to the industrial dexterous hand Linker Hand O6; under the two companies' February agreement for a total procurement of 100,000 units, over 30% of the order has now been fulfilled source

· QJ Robotics whole-hand tactile 2.0: extends tactile sensing from the fingertips across the finger pads and palm, with fingertip marker density reaching 125 points/cm², combined-force precision of 0.03N, and a refresh rate of 120Hz — 4x the previous generation; the flexible e-skin offers force resolution of 0.01N with a range up to 200N; a hub board in the palm can connect 5 sensor channels and packs up to 6 TOPS of on-device compute, cutting the host's perception compute load by up to 90% source

· Robot Components Industry Development Alliance: launched by JD Industrials together with Schaeffler and others at the 2026 World Robot Conference, aiming to use real industrial use cases to drive next-generation component performance, lifespan, and safety improvements, and to bring costs down through standardization and scale. Schaeffler Greater China e-drive division president Chen Xiangbin said the alliance "bridges the gap between components and actual robot deployment" source

FutureX · Physical AI Daily — Issue 106 (09/01)

Shawn — Mon, 31 Aug 2026 14:50:25 +0000

Today's Highlights

· DiDi's next-generation Robotaxi R2 launches fully driverless paid trial rides in Beijing and Guangzhou

· Galbot (Chinese embodied-AI startup) opens three fully autonomous robot retail stores in Hong Kong, doors open September 1

· FedEx scales up Dexterity's truck-loading robots to Hagerstown hub operation

· AgiBot (Chinese humanoid startup) open-sources AGIBOT WORLD phase three, first batch of 11,430 real-robot reinforcement-learning trajectories

· Zhongda Leader (Chinese reducer maker)'s profit eaten by price war, net profit attributable to parent down 40.09% year-on-year

· Universal Robots' parent company sues Jaka (Chinese cobot maker)'s German subsidiary at the Unified Patent Court in Europe

· Momenta posts RMB 1.60 billion in first-half revenue, adjusted net loss narrows 97%

Paper Progress

VLAct: continued pretraining for VLA through representation, not more data · vla

Robot trajectory collection costs cap dataset scale, so this paper treats representation quality as the bottleneck instead, running multi-embodiment continued pretraining under a fixed data budget before task fine-tuning. It reaches 82.6% on LIBERO-Plus and 92.5% on RoboTwin 2.0, surpassing industrial VLA systems such as ABot-M0 and LingBot-VLA; on the unseen humanoid embodiment RoboCasa-GR1, it beats GR00T-N1.6 trained on the full dataset while using only 20% of the downstream trajectories. The whole pipeline uses open-source data, and training ran on just 16 GPUs. Hugging Face community interest: 49↑.

Senqiao Yang et al. · arXiv 2608.27550 source

How large a video diffusion model can 5,500 hours of driving data support · world-model

Driving data is too costly to buy or scrape at scale, which forces scaling laws to be recalculated. valeo.ai trained an entire family of models from scratch, ranging from 1M to 9B parameters, and found that validation loss follows a power law with respect to both model size and training exposure, but degrades much faster with exposure — meaning that with limited compute, extending training is more cost-effective than scaling up the model. The resulting 9B model sets a new open-source SOTA for driving video generation on nuScenes; code and weights have been released.

Victor Besnier et al. (valeo.ai) · arXiv 2608.28404 source

MAGP: giving geometric reconstruction a real-world scale · perception

Existing reconstruction methods only produce relative geometry, with scale drifting across scenes, viewpoints, and input configurations — object sizes and spatial distances don't match the real metric scale that robot actions operate in. MAGP uses scale-equivariant augmentation to push the model to recover metric geometry from camera parameters and depth observations, cutting absolute error from 2.01 meters to 0.07 meters on ETH3D, MegaDepth, and ScanNet++. Plugged into multiple robot policies, it boosts RoboTwin performance by up to 6.26%.

Fengjun Zhong et al. · arXiv 2608.27497 source

Aero Hand Open: a tendon-driven hand open-sourced together with its simulation model · manipulation

Tendon-driven hands are cheap — moving the motors out of the joints allows for a smaller hand where a single cable can drive several joints — but the tradeoff is that simulation is hard to build and coupled joints can't be commanded independently. This hand writes the cable transmission itself into the simulation model, paired with an identified bidirectional actuation mapping (including three-way coupling for the thumb) and an RL training package. Policies trained purely in simulation transfer directly to the real hand, with no fine-tuning or state estimation needed. The mechanical design, simulation, mapping, training environment, and deployment stack are all released.

Nan Wang et al. · arXiv 2608.28578 source

Walking while seated: a humanoid robot learns omnidirectional locomotion on a wheeled chair · locomotion

A near-direct-drive humanoid standing upright must continuously output joint torque, whereas sitting shifts the body weight onto a chair. The authors have the robot learn omnidirectional locomotion on a passively rolling chair, with the pelvis and seat left unfixed, propelled by intermittent foot pushes against the ground. The policy uses no motion-imitation reward — the actor only takes proprioception and velocity commands as input — and the best seated policy actually outperforms a standing policy on velocity tracking, transferring zero-shot to a real Unitree G1.

Kango Yanagida et al. · arXiv 2608.28090 source

Using contact rewards to guide exploration, a quadruped mobile manipulation robot learns to move a chair · manipulation

Non-grasping manipulation like pushing, nudging, and prying gets stuck on hybrid dynamics and sparse contact. This paper adds a dedicated exploration critic to multi-critic RL, using a dense "find contact" reward to steer the end-effector toward meaningful contact points, then gradually decaying its weight back to the task-optimal policy; candidate contact points come directly from a general-purpose grasping algorithm, which lets the approach generalize across object shapes. Across three tasks — pushing a box, moving a chair, and opening a dishwasher — the chair-moving policy was deployed on a real quadruped mobile manipulation robot.

Simone Tolomei et al. · arXiv 2608.28140 source

CoCoBench: breaking apart multi-agent collaboration failures · benchmark

Overall success rate masks duplicated effort, sequencing violations, resource contention, and handoff desynchronization all at once. CoCoBench uses 897 oracle-verified household task instances to score four collaboration components separately: task allocation, sequencing, mutual exclusion, and handoff. Across 11 mainstream MLLMs evaluated, collaboration ability turns out to be highly component-specific: a model that leads on overall score can still fail badly on one particular collaboration dimension.

Yang Chen et al. · arXiv 2608.28266 source

What happens when a robot mishears a single word · benchmark

The risk introduced by voice interfaces is rarely tested in isolation. The authors overlay simulated ASR recognition errors onto SafeAgentBench and POEX and find that some errors preserve semantic structure while amplifying harmful ambiguity, and others directly weaken the model's refusal behavior, letting unsafe plans get generated and executed. Automatic error correction sometimes suppresses the risk, but not reliably.

Sihan Jia, Oliver Lemon · arXiv 2608.28518 source

Other papers today: DeicticVLA unifies language and deictic gesture instructions into a single VLA (arXiv 2608.28108 source); CAVE-NAV does 3D autonomous navigation of underwater caves with a VLM (arXiv 2608.27793 source); ChainSplat learns deformation dynamics of cables, ropes, and hoses from multi-view RGB video (arXiv 2608.28570 source); PanelShield adds verifiable closed-loop safety planning for industrial panel operation (arXiv 2608.28305 source); a year of running through subarctic forests, with the authors reviewing the practical difficulties of autonomous navigation under canopy cover (arXiv 2608.27628 source); Picking Bins Empty uses online self-learned grasp points to solve deadlocks in emptying industrial bins (arXiv 2608.28175 source); AcrossVAM1.0 splits robot video prediction into object-level motion and dense appearance streams (arXiv 2608.28491 source); PHR-VLA adds planning-horizon reasoning to VLA (arXiv 2608.27609 source).

Open Source · Tools · Benchmarks

· AGIBOT WORLD phase three: AgiBot has released a reinforcement-learning dataset, with a first batch of 11,430 real-robot interaction trajectories covering industrial and household scenes across 14 categories of interaction tasks, including plugging in network cables and unlocking doors with a key. The data was collected by the next-generation Spirit G2 platform and will be released in five phases, with digital-twin simulation data released in parallel. Real-robot RL trajectories have long been scarce in public datasets, and this batch is large enough to support third-party reproduction. source

· Microduck: a $399 bipedal robot duck jointly released by Hugging Face and its French subsidiary Pollen Robotics — 25 cm tall, under 800 grams, packing 15 servos, a wide-angle camera, a small lidar, and dual IMUs. The full-stack open-source release is the selling point: 4,096 virtual ducks can be trained in parallel in MuJoCo for RL, producing a new gait in roughly one to two hours, then exported as ONNX and deployed to the physical unit. The device is manufactured by Shenzhen Seeed Studio (Chinese hardware contract manufacturer). source

· Berkeley Humanoid Lite: an open-source humanoid reference design from UC Berkeley — 0.8 meters tall, about 16 kg, with full hardware priced under $5,000 at US market rates. The actuators use brushless DC motors paired with 3D-printed cycloidal reducers and position encoders; the housing and gears are designed to fit within a 200×200×200 mm desktop-printer build volume, with the larger 6512 actuator listed at about $188. source

Funding & Deals

NVIDIA × Hugging Face｜Acquisition｜$12.9 billion · adjacent ⚠️ unconfirmed report

The Information reported on August 27 that NVIDIA has agreed to acquire the open-source model hosting platform Hugging Face, with CNBC, Reuters, and Forbes following up subsequently; neither company has officially confirmed the deal, and reports say a final agreement has not been signed and could still change. On the embodied-AI side, Hugging Face acquired France's Pollen Robotics last year and just released Microduck on August 27, so ownership of the LeRobot ecosystem would change hands along with the deal. If completed, this would be NVIDIA's largest acquisition ever, roughly twice the scale of its 2020 acquisition of Mellanox.Source: The Information / CNBC

Shunxing Qidian (Hangzhou)｜Seed round｜tens of millions of RMB · hardware ⚠️ compiled from public filings

Investors include two industry players, Galbot (Chinese embodied-AI robotics company) and Zhongke Diwuji, alongside Bamboo Capital (Chinese VC), Hangzhou Kechuang Group, Zhaofeng Co., and Duowei Capital. The company positions itself as a one-stop service provider for embodied robotics, handling the full chain for OEMs from core components and joint modules to full mass-production delivery, without building its own consumer-facing brand. As integrator OEMs have piled in, contract manufacturing and mass-production delivery are starting to attract dedicated funding of their own. Data compiled by Securities Star based on Tianyancha's public disclosure filed August 30.Source: Sina Finance source

ABLE Labs (South Korea)｜Series A｜about $8.7 million (KRW 12.1 billion) · adjacent

Led by Innopolis Partners, with participation from the Korea Development Bank, Capstone Partners, Horizon Investment, Magna Investment, Futureplay, and NAU IB Capital. The company builds lab automation robots and plans to expand from liquid-handling robots into autonomous laboratories, while also using the experimental data infrastructure it builds to train models. 2025 revenue was KRW 1.29 billion, up 113% year-on-year, and first-half 2026 order backlog already exceeds full-year 2025 revenue by more than double; customers include KAIST, Pusan National University, and the Korea Research Institute of Chemical Technology, and the company is also co-building an organoid drug evaluation platform with Johns Hopkins University.Source: WOWTALE source

Xuanchuang Robotics｜Series A｜tens of millions of RMB · embodied

Following the closing of a Series A1 round, this round's investors have been revealed as Shenzhen Qianhai Fangzhou Capital, Guangyang Co., and others. The company recently won an order for more than 100 units from PetroChina, with deliveries running at 20 units per month.Source: DoNews source

Commercialization & Deployment

DiDi's next-generation Robotaxi R2 launches fully driverless paid trial rides in Beijing and Guangzhou · autonomy

R2 was co-developed with GAC Aion and purpose-built for Robotaxi service, running DiDi Autonomous Driving's full-stack L4 software and hardware, equipped with 33 sensors and a tri-domain fused central computing platform, plus multiple layers of redundancy, with safety standards aligned to C-NCAP and Euro NCAP five-star ratings. After delivery in January this year, the vehicle first ran public-road testing in Guangzhou, Beijing, Shenzhen, and elsewhere; it can now be booked via the DiDi app in designated demonstration zones in the two cities. The cabin features a 17.3-inch ceiling-mounted screen, and voice commands alone can complete PIN verification, start the ride, adjust the air conditioning, and open windows. Just the day before, Caocao Mobility pushed RoboX into Abu Dhabi — Chinese Robotaxi fleets are simultaneously moving toward removing safety drivers and toward overseas markets.Source: PR Newswire source

Galbot opens three fully autonomous robot retail stores in Hong Kong, doors open September 1 · embodied

One store each in Hung Hom, Kai Tak, and Wan Chai — Galbot's first stores outside mainland China. Founder and CTO Wang He said at the launch ceremony that this step moves the company from R&D into actual operations and is a key step toward global markets, and that he hopes to build embodied-AI labs jointly with Hong Kong universities and local industry partners. Galbot has already opened roughly 200 stores across nearly 50 cities in mainland China. Financial Secretary Paul Chan attended and cited Hong Kong's dense population, diverse retail environment, and high degree of internationalization; Hong Kong Investment Corporation has been a Galbot shareholder since 2024. What retail stores demand of robots isn't motion difficulty but continuous availability while unattended — three stores opening on the same day is about density, not a single showcase.Source: China Daily source

FedEx scales Dexterity's truck-loading robots from pilot into the Hagerstown hub · industrial

The deployment pairs Dexterity's Foresight world model with its Mech loading robot, at the Hagerstown Hub in Maryland, at a scale considerably larger than the earlier pilot. Mech is a dual-arm system built narrow enough to work inside trailers; Foresight combines vision, depth, and touch to predict how actions affect the physical world, and loading requires weighing space utilization, load stability, and speed across three spatial dimensions plus time. Trailer loading is one of the most physically demanding steps in package logistics that traditional automation has long failed to crack, and FedEx loads tens of thousands of trailers a day across the US. Kawal Preet, FedEx's Executive Vice President of Planning, Engineering, and Transformation, said the expanded partnership is about "evaluating and deploying physical AI in real-world environments." Dexterity founder and CEO Samir Menon called Hagerstown a template for replicating across the network.Source: AI Insider source

Momenta posts RMB 1.60 billion in first-half revenue, adjusted net loss narrows 97% · autonomy

The first half-year report since its Hong Kong listing: revenue of RMB 1.60 billion, up 76% year-on-year, gross profit excluding share-based compensation of RMB 1.21 billion at a 75% gross margin, with adjusted net loss narrowing to RMB 14.097 million. Revenue splits into two segments — technology development services at RMB 995 million (62.1%) and licensing services at RMB 607 million (37.9%); licensing revenue was only about RMB 23 million back in 2023. As of end-June, the company had 105 mass-produced vehicle models and 219 secured design wins cumulatively, with vehicles running its system now exceeding 1.10 million units. R&D spending was RMB 1.16 billion, or 72.6% of revenue. Robotaxi is already being tested in Shanghai, Suzhou, Abu Dhabi, and Munich, and in July the company obtained Germany's KBA nationwide urban road L4 testing permit; Momenta told National Business Daily that its first mass-produced Robotaxi model will launch in Q4 2026, running the R7 world model, with several hundred vehicles deployed domestically and internationally by year-end. With L2 penetration already past 70% and continuing to spread into lower-tier markets, whether the 75% gross margin can withstand the price war is the real thing to watch in the second half.Source: National Business Daily source

MEgo Technology's 20,000th MEgo capture device rolls off the line, embodiment-free data surpasses 1 million hours · adjacent

The 20,000th MEgo unit was delivered by founder and CEO Yao Maoqing to JD.com, going directly into use in real business scenarios; the same day, MEgo announced an embodiment-free data business partnership with Tencent Robotics X Lab. The company says cumulative embodiment-free data production has surpassed 1 million hours, all collected in real open environments, covering 22 major scene categories, over 10,000 real environments, over 50,000 object categories, and more than 500 fine-grained tasks. MEgo (Chinese motion-capture-data startup) was incubated inside AgiBot before spinning off independently, and has raised hundreds of millions of RMB, with investors including China Telecom, HongShan (formerly Sequoia China), and Baidu Ventures. Now that capture devices have scaled to 20,000 units, the competitive front on the data-supply side is shifting from device count to data structure and scene coverage.Source: The Beijing News source

Hai Robotics' HaiClimber surpasses 10,000 units contracted worldwide · industrial ⚠️ vendor-reported

Since its global launch in 2025, HaiPick Climb has been selected for customer projects in 12 countries across Asia, Europe, North America, South America, and Australia, with customers including Anta, Panasonic, Arvato, Itochu, Mettler Toledo, and True Protein. Founder and CEO Rocky Chen said what's more notable than the 10,000-unit figure is that 8 customers have expanded existing systems or repeat-purchased for new projects. Two different framings appear in the same release: the body text says "contracted for use in customer projects," while Americas Sales VP Hunter Senn's quote says "deployed" — and there is typically a delivery lag of several months between a contract and installation.Source: Construction Business News ME source

Industry Trends

Universal Robots' parent company sues Jaka at the Unified Patent Court in Europe · industrial

Teradyne Robotics has sued the German subsidiary of Jaka Robotics at the Unified Patent Court (UPC) in Copenhagen, alleging infringement of both software and hardware patents, covering a broad range of collaborative robot models that Jaka sells in the EU. The patents at issue are held by the Danish subsidiary Teradyne A/S, tied to the Universal Robots A/S business unit. Teradyne Robotics Group President Jean-Pierre Hathout said the company would enforce its patent and IP rights against what it considers "unlawful and improper imitation." This is Teradyne Robotics' second IP lawsuit in Europe in 2026, following an earlier copyright suit in Germany over Universal Robots software against another robotics maker's subsidiary. The ruling can apply across 17 of the UPC's current 18 member states, with non-UPC countries such as the UK and Spain also drawn into the proceedings. This is the first time Chinese cobot makers' expanding European market share has run into a pan-European patent enforcement tool.Source: All About Industries source

Zhongda Leader's half-year report: revenue up 15.6%, net profit attributable to parent down 40% · hardware

This manufacturer based in Cixi, Zhejiang, mass-produces all three types of reducers — planetary, RV, and harmonic. First-half revenue was RMB 597 million (+15.61%), net profit attributable to parent RMB 27.78 million (-40.09%), non-recurring-adjusted net profit RMB 19.51 million (-48.15%), with overall gross margin at 23.28%, down 4.46 percentage points year-on-year, and operating cash flow turning negative to -RMB 56.49 million. The revenue mix has shifted: smart actuator units brought in RMB 253 million (42.40%), becoming the largest revenue source for the first time; precision reducers brought in RMB 134 million (22.39%), also the single biggest drag on profit this period. Per pricing cited by Robot Vision Forum (Chinese robotics industry outlet) from public industry information, harmonic reducer unit prices have fallen from over RMB 3,000 during the era of import monopoly to the low-thousands range, with some models down near RMB 800; China's harmonic reducer annual production capacity is heading toward a million units, while global humanoid shipments in the first half were still only in the tens of thousands, with each robot using 8 to 14 harmonic reducers. The same report notes that Leaderdrive (Chinese harmonic reducer maker) still holds gross margin above 30% on the strength of high-end products and top-tier design wins. Humanoid shipment volumes ramping up and component makers actually turning a profit are, for now, two separate things.Source: Robot Vision Forum source

Li Auto unveils next-generation autonomous driving architecture MindVLA-o1 at GTC · autonomy

Zhan Kun, head of foundation models, offered the diagnosis that today's end-to-end systems are all essentially "learning to drive by watching 2D video" — BEV flattens the world and loses height, while OCC is three-dimensional but lacks semantics. MindVLA-o1 is built around a native multimodal MoE Transformer, using a native 3D ViT encoder that unifies 3D geometric and semantic understanding in a single encoding pass, reducing lidar to a "ruler" that geometrically calibrates vision. Li Xiang (Li Auto founder and CEO) said unified modeling stably extends perception and reasoning to over 500 meters. On the behavior side, VLA-MoE with parallel decoding is paired with discrete diffusion for multi-round trajectory refinement; on the training side, step-by-step reconstruction has been replaced with feed-forward scene reconstruction, nearly doubling rendering speed and cutting overall training cost by about 75%. 3D ViT is compute-hungry, and deployment relies on Li Auto's own Mach 100 chip at 1,280 TOPS per chip, first fitted to the all-new L9. The team evaluated nearly 2,000 model architecture configurations.Source: Autohome source

Qiyuan and Zhijie both disclose L3 approval progress the same day, both still awaiting ministry-level review · autonomy ⚠️ vendor-reported

Xuan Wei, product general manager at Qiyuan Auto, said on social media that the Qiyuan GT7 already meets the design standards of China's national L3-level autonomous driving architecture and has completed L3 access testing and filing, with further rollout still pending review by national ministries and related regulations. He described the L3 threshold as requiring that the eight systems — power supply, compute, perception, positioning, communication, steering, braking, and interaction — must never all go offline simultaneously in the event of component failure, and the vehicle must be able to safely pull over. Zhijie's framing goes one step further: product manager Li Xinyue said in a livestream that the Zhijie RX has been undergoing L3-related autonomous driving testing in Shanghai, Wuhu, and elsewhere, though aggregator headlines rendered this as "access testing," which differs from the body text. Richard Yu (Huawei executive) said the Zhijie RX is the first coupe SUV built for L3, with L3-architecture-equipped orders accounting for over 90% of pre-sales at an average price of RMB 380,000. Both companies remain at the filing-and-testing stage, and whether the feature can be activated depends on regulators, not the vehicles themselves.Source: PChome source, Sina Auto source

BYD and PaXini upgrade from an investment relationship to strategic cooperation, opening vehicle plants to robots · embodied ⚠️ vendor-reported

The agreement was signed August 25 at BYD's global headquarters in Shenzhen and announced by BYD on the 31st, covering both data collection and robotics. BYD contributes complex, exacting vehicle- and component-manufacturing processes, extensive factory environments, and supply-chain resources, giving robots access to real industrial test settings and generating operational data usable for training embodied models; PaXini (Chinese tactile-sensing robotics startup) contributes tactile sensors, multi-dimensional perception, and multimodal embodied data collection. Founded in Shenzhen in 2021, the company builds high-precision tactile sensors, dexterous hands, and the Tora series of humanoids; it raised over RMB 1 billion in March at a valuation above RMB 10 billion, with JD.com also among its shareholders. BYD invested in PaXini in April 2025 and had previously invested in AgiBot. PaXini has previously partnered with Geely, with its Tora-One performing wire-harness assembly and glue application at the Qiantang plant; the company's stated figures at the time were 8 hours of continuous fault-free operation and a 99.8% task success rate — figures for a single demonstration run, not the same thing as sustained production-line operation.Source: CnEVPost source

California Teamsters sue over the legislative process behind new driverless trucking rules · autonomy

California's driverless trucking framework, drafted since 2021, took effect April 28, requiring 1 million miles of testing before commercial freight operation. Teamsters California filed suit in Alameda County Superior Court on August 5, not challenging the substance of the rule but only the process — alleging that the DMV pushed the rule through a fast-track process meant for regulations with first-year costs or benefits under $50 million, "evading the law's requirement that the agency study and publicly disclose the economic impact." The lawsuit hasn't frozen approvals — the permit queue continues to move, with 28 permit holders including Aurora, Kodiak AI, Gatik, and Plus. Plus's self-published readiness scorecard on August 10 showed a 93.4% safety case, 99.8% autonomous-driving mileage, and 85.2% of trips with no remote assistance, with a target of removing the driver by 2027. Its commercial pilot isn't in California but on a 600-mile daily route along the I-35 corridor in Texas, run jointly with International Motors and Ryder. What's constraining these companies for now is the calendar, not the courts.Source: FreightWaves source

Chinese robotic lawnmower makers turn en masse toward Europe · adjacent

According to IDC data, global robotic lawnmower shipments grew 63.8% year-on-year in 2025 to 1.992 million units, with the top six all Chinese manufacturers. Ninebot, Roborock, and a cluster of startups are preparing for IFA Berlin, opening September 4. Shenzhen-based LDROBOT (Anthbot brand) moved from being a lidar supplier into making complete units; co-founder and CEO Gavin Guo told Nikkei that lawnmower revenue more than doubled year-on-year to RMB 200 million in the first half of this year, now accounting for 38% of total revenue, with a target of reaching the top three within three years. Also Shenzhen-based Mova (TerraMow brand) uses a triple-camera approach, bypassing lidar and GPS to bring prices down to €669–€1,149; founder Harry Zhang, a DJI alumnus, says his reasoning is that users skew older and won't tolerate configuring a device via smartphone. There is almost no lawn-care demand within China itself, so this product line has been a pure export business from day one, and US import restrictions have pushed essentially all the growth toward Europe.Source: Nikkei Asia source

GXO's CEO: warehouses need better robot hands, not humanoids that can run and jump · industrial

GXO, a logistics company with $13.6 billion in annual revenue, is running pilots with five humanoid robot suppliers and will add a European pilot this year. CEO Patrick Kelleher said the record-setting sprints and high jumps at the World Humanoid Robot Games in Beijing were of no use to him — what he wants is a robot that can unload trucks and also sort lipstick. He's focused on what he calls the "single-digit problem," meaning getting robot hands closer to human hands: "One important advance we've been involved in this year is getting a humanoid's hand to move like this, which requires multiple flexible joints and will change the range of tasks a humanoid can do." Industrial robots typically have four to six degrees of freedom in their motions, whereas humanoids can now exceed 100. He stressed this isn't about replacing the existing 150,000 employees: "Our warehouse environments have 25% turnover, and we can't hire enough people as it is." GXO is simultaneously developing fully autonomous forklifts and humanoids that can work in cold-storage facilities at minus 32 degrees.Source: Fortune source

Hardware · Supply Chain

· Qualcomm Dragonwing IQ10: Anshuman Saxena, Qualcomm VP and GM of ADAS and Robotics, said the platform has shipped to multiple customers since its launch earlier this year, with several Dragonwing-based products now in mass production in China, spanning AMRs to humanoids; publicly disclosed partners include Figure AI and NEURA Robotics, and the company has also set up a robotics center in Japan. He pushed back on the "single big brain" approach, arguing intelligence should be distributed to wherever the actuators are. source

· XenseVision Prima 1: a purely direct-drive dexterous hand with 22 active degrees of freedom and five-finger visuo-tactile sensing; at a trade show it drew building outlines with a paintbrush, adjusting grip force in real time based on friction between the brush and paper, aimed mainly at academic research and model training. Alongside it, the Flex 2 is a 23-DOF hybrid-drive variant with air cooling, aimed at precision assembly and industrial sorting. Founder and CEO Xia Yuxuan said the two lines are complementary, not a replacement for one another. source

· Robot Component Industry Development Alliance: launched by JD Industrials during WRC 2026, with Schaeffler Greater China e-mobility division president Chen Xiangbin attending the launch ceremony. The alliance spans transmission, motors, sensing, motor control, end effectors, whole-machine integration, and precision testing, aiming to use real industrial applications to set lifespan and safety specifications for next-generation components. source

FutureX · Physical AI Daily — Issue 105 (08/31)

Shawn — Sun, 30 Aug 2026 14:50:50 +0000

Today's Highlights

· AgiBot (Chinese humanoid robotics startup)'s humanoid robot factory built with Minth Group in Serbia begins production, targeting over 5,000 units a year

· Sharpa's dual-arm robot enters a Shanghai DQ store, making a Blizzard in 55 steps

· Pony.ai signs South Korea's FutureLink, plans to deploy 200 L4 vehicles in Seoul, with certification and commercialization timeline still undecided

· China's industrial robot output reached 635,000 units in the first seven months, up 28.5% year-on-year

· Geek+ (Chinese warehouse robotics company)'s new orders rose 35.5%, while its stock price fell nearly 40% below its IPO price

· An OpenAI executive puts affordable L5 at 5 to 10 years away; Musk replied with a clown emoji

Research Progress

Open Source · Tools · Benchmarks

· Bifrost Manifold: This policy evaluation platform for robotics researchers is now open for early-access applications. It can run thousands of rollouts simultaneously on GPUs — a LIBERO sweep that used to take 8 hours now finishes in 30 minutes⚠️ vendor claim. The platform's agent also automatically clusters failure cases and ranks them by impact; models like OpenVLA, GR00T, π0, and Octo can all be evaluated on LIBERO, RoboCasa, RoboMimic, and CALVIN, with replay footage rendered as 3D visualizations via Rerun. (Bifrost AI)

· Microduck: Hugging Face and Pollen Robotics have jointly opened preorders at $399. The bipedal little robot stands just 25 cm tall and weighs under 800 grams, packed with 15 motors, a camera, and a small lidar unit, paired with a 2600mAh battery good for about an hour of runtime; it will ship before Christmas. It doesn't speak, interacting only through sound, and is positioned as an open-source desktop robot with reinforcement-learning-trained posture control. (Engadget, TechCrunch)

· Model Hardware Standard: Anthropic has released a shared specification allowing AI agents to safely operate physical devices, now open for research preview (previously reported). source

Funding & Deals

LatentVerse (潜界科技) ｜ Seed round ｜ hundreds of millions of RMB · world-model

The list of investors in this round has finally been revealed: Hillhouse Ventures, Clearvue Partners, and Innoangel Fund are all on it, alongside two leading embodied-AI companies, AgiBot and Robot Era (Chinese humanoid startup). The round was first disclosed on August 11, and the company itself is less than two months old, with a core team drawn from Tsinghua University's Institute for Interdisciplinary Information Sciences. Founder Hu Yucheng is a student of Robot Era founder Chen Jianyu. The core model they're developing is called UTAM — Unified Tactile-Action Model — which folds intent understanding, world-state prediction, robot action generation, and tactile feedback all into a single embodied-native foundation model trained together. Less than two months elapsed from the start of fundraising to closing, and several VCs have already secured spots in the next round.Source: China Venture source

Psi Robot (灵初智能) ｜ New round ｜ over $100 million · embodied

The investors behind this week's disclosed nine-figure-dollar round have now been named: Tuopu Group, Chery Holding's Ruicheng Fund, Lens Technology, 37 Interactive Entertainment, Wuhu Municipal Investment Holding Group, and Fosun Chuangfu are all in, with existing shareholder Zhuhai Science & Technology Industry Group adding significantly to its stake. The company's proprietary Psi-SynEngine embodied data-collection engine has already been validated at small scale in logistics customers' warehouses. Following this round, core investment will shift more toward world-model research and development.Source: Cailianshe VC Desk source

Sharpa ｜ Cumulative funding ｜ over RMB 4.5 billion ｜ post-money valuation RMB 22 billion · embodied

The post-money valuation figure has now been confirmed at RMB 22 billion. Corporate backers include Alibaba, Meituan, Tencent, JD.com, and Transsion, alongside institutional investors Sequoia China, Qiming Venture Partners, Meituan Longzhu, and Guanghe Ventures. The company was founded in late 2024 by the three founders of Hesai Technology (Chinese lidar maker) and had never previously disclosed its fundraising history.Source: STCN source

Weekly private-market roundup ｜ Aug 22–28 ｜ disclosed funding totaled approx. RMB 16.15 billion · adjacent

According to Cailianshe VC Desk's tally, 136 funding events occurred this week, with advanced manufacturing accounting for the largest disclosed total at approximately RMB 6.625 billion. XPeng Robotics' first-round raise of over $900 million was the single largest deal of the week (previously reported). By round stage, seed/angel rounds led in count with 49 deals, about 36% of the total; Series A rounds, though only 36 in number, accounted for the highest disclosed total, roughly RMB 8.215 billion. By region, Shanghai led with 31 companies funded, followed closely by Shenzhen with 26.Source: Cailianshe VC Desk source

Commercialization & Deployment

AgiBot and Minth's humanoid robot factory begins production in Šabac, Serbia

On August 29, Serbian President Aleksandar Vučić personally toured the production floor in Šabac and, in a lighthearted moment, played goalkeeper against a shooting robot. The factory was built through a partnership between AgiBot Innovation, Minth Group, and Serbian partners, and is the first robot factory built in Europe through Chinese corporate investment. In his remarks, Vučić thanked China and noted that Minth has already created nearly 5,000 jobs in Serbia, saying the country "is becoming a pioneer in Europe's scaled mass production of humanoid robots." He continued: "We hope that in the future, this factory will be able to assemble and produce more than 5,000 humanoid robots a year." These robots are intended for use in healthcare, agriculture, industry, and defense. The first batch, rolling off the line in 2026, will initially be deployed at the 2027 Belgrade Specialized Expo, with full-scale mass production not expected until 2027. The figure of 5,000 units a year remains an expectation stated by the president, distinct from the production line's current actual capacity.Source: Guancha source

Sharpa's dual-arm robot goes to work at a Shanghai DQ store, operating 12 hours a day

From cup retrieval to mixing to inverted delivery, the entire 55-step process runs without any human intervention. Starting August 29, a white dual-arm robot named "Xiao Sha" has been working at a DQ store on Wujiang Road in Shanghai, taking about 6 minutes to make a Blizzard, compared to two to three minutes for an experienced human staffer. The store operates from 10:00 to 22:00 year-round, with the robot counter and the human counter running as two parallel lines; online reservation slots are already booked out several weeks in advance. Sharpa's first principle for the project was never to modify the environment: the equipment, ingredients, processes, and tools all remain exactly as they were, with the robot handling the same paper cups and mixing spindles used by human staff. The company developed a proprietary hierarchical end-to-end model called CraftNet that incorporates tactile input, paired with the Sharpa Wave dexterous hand; 98% of the 55 steps directly rely on tactile sensing — for instance, sensing friction when pulling a cup, or continuously adjusting grip force in response to resistance during high-speed mixing. Co-founder Li Yifan acknowledged that the first-generation robot has not yet turned a profit, and that this early wave of deployed scenarios industry-wide has very little chance of positive ROI in the short term. He cited three long-term variables: whether the robot's operation can approach human-level performance, whether the hardware can last more than 10 years, and whether mass-production costs can drop from the hundreds of thousands of RMB range to around RMB 100,000.Source: STCN source

Pony.ai signs FutureLink, to deploy 200 L4 vehicles in Seoul by 2028

An initial batch of 10 vehicles will go through South Korea's certification process first, with the remaining 190 to follow once certification is complete. On August 28, Pony.ai signed a partnership in Seoul with South Korean mobility company FutureLink. The vehicle is based on BAIC's Arcfox Alpha, with the seventh-generation autonomous driving system integrated directly on the vehicle production line rather than retrofitted afterward, and redundancy built into critical functions like braking, steering, and power. Service will initially launch in Seoul's Gangnam autonomous driving pilot zone, before expanding to the rest of Seoul, the greater metropolitan area, and other major cities; specific certification and commercialization timelines have not yet been set, and the two companies are also considering forming a joint venture. FutureLink said it has already conducted day-and-night testing in Gangnam using a Kona Electric fitted with Pony.ai's kit, covering roughly 80,000 km without incident. Founder Peng Jun told The JoongAng Daily that South Korea was chosen because of its large automotive industry scale, government willingness to improve regulation, and relatively high taxi fares: "Seoul's high urban density and complex road conditions make it one of the more technically demanding regions for autonomous driving — we see South Korea as a proving ground for the technology." He also noted that operations in Guangzhou and Shenzhen are now profitable. The production cost of the seventh-generation kit is 70% lower than the previous generation, and the company plans to cut costs by a further 20% in 2026, aiming to bring the combined cost of kit and vehicle below RMB 230,000 by mid-2027. In Shenzhen, the fleet averaged 23 paid rides per vehicle per day this past February.Source: The Korea Times source, The JoongAng Daily source

South Korea's Jeonbuk National University builds a physical AI pilot lab to add "autonomous judgment" to automated production lines

The 846-square-meter lab houses 9 robot cells and was built last year through a KRW 6.77 billion proof-of-concept project. A robotic arm places steering wheels onto inter-process transfer equipment, another arm continues processing and assembly, and AMRs shuttle materials between cells. When a reporter accidentally stepped into a robot cell's safety zone, an alarm sounded immediately and the equipment stopped, with the digital twin display on the wall syncing the updated status about 0.5 seconds later. The lab has separate process configurations for three product types — steering wheels, mirrors, and others — with each cell assigned its own agent responsible for that process, plus an upper-level integration agent overseeing the whole system. When sensor readings go out of range, the cell-level agent flags the anomaly while the upper-level agent assesses its scope and offers recommendations to workers. Currently the system only handles anomaly detection and response suggestions — actually adjusting equipment or resuming a process still requires the main control system. In other words, if a piece of equipment fails and the issue touches process chains involving other facilities or logistics robots, human or upper-level control system intervention is still required. Kim Sun-tae, head of Jeonbuk National University's Physical AI Convergence Technology Program, said: "Putting humanoid robots into the mix won't immediately raise factory productivity." In his view, what really matters is a factory operating system that lets existing equipment coordinate autonomously — not the robots' movements themselves.Source: Digital Today source

Singapore Post trials autonomous logistics vehicle in Tampines, running only 4 to 7 a.m.

The vehicle being trialed is the Zelos Z10, shuttling between SingPost's regional e-commerce logistics hub and the T-Space building on Tampines Concourse Link, equipped with 360-degree sensors, though a safety operator still rides along on each trip. The news was announced late on August 29 via Facebook by Ma Yam Keng, Mayor of North East District and MP for Tampines GRC. The trial hours are timed to avoid daytime traffic peaks entirely.Source: 8world source

A 2.3-mile Tesla Robotaxi trip is quoted at 66 minutes

A passenger identifying as a "Tesla Ambassador," Reggie Overton, hailed a ride in Austin for a 2.3-mile trip, only for the system to estimate 1 hour and 6 minutes. He initially thought it was a software bug and called customer service; after several rounds of back-and-forth he was told the route couldn't be recalculated, and he ultimately had to change his destination just to end the ride, then hailed a separate Uber. Tesla later refunded him. He wrote on X: "For a 2.3-mile trip, this experience was extremely frustrating, especially since customer service couldn't correct an obvious routing problem." Tesla's Robotaxi service currently remains limited to Austin, Dallas, and Houston in Texas, and Miami, Orlando, and Tampa in Florida. The invite-only launch event for the Cybercab is set for September 3. Musk said months ago that he expects robotaxi service to be available to about half the U.S. population by year-end.Source: Chron source

Industry Developments

China's industrial robot output hits 635,000 units in first seven months, up 28.5% year-on-year

According to data from the China Machinery Industry Federation, cumulative industrial robot output for the first seven months reached 635,000 units, up 28.5% year-on-year, while cumulative power generation equipment output over the same period reached 208.97 million kW, up 2.8%. Value-added output grew year-on-year across all five major machinery-industry categories: instrumentation manufacturing grew 9.4%, specialized equipment manufacturing 8.8%, general equipment manufacturing 7.9%, automotive manufacturing 7.2%, and electrical machinery and equipment manufacturing 6.5%.Source: People's Daily source

Geek+'s new orders rise 35.5% while stock falls below IPO price

New orders in the first half reached RMB 2.385 billion, up 35.5% year-on-year, nearly 1.9 times revenue for the same period; revenue came in at RMB 1.284 billion, up 25.3%, with gross margin rising from 35.1% to 35.8%. Revenue from outside mainland China accounted for over 75% of the total, with overseas sales carrying a gross margin of 46.2%. Subscription-service orders reached RMB 156 million, up over 75% year-on-year, with the Americas region up as much as 455%; orders in manufacturing scenarios grew over 600%, while the pallet-to-person product line grew over 200%. Adjusted net loss was RMB 60.6 million, narrowing 32.1% year-on-year; the company also noted that excluding RMB 44.5 million in embodied-AI R&D spending, the adjusted loss would have narrowed by 81.9%. However, the reported net loss widened from RMB 50 million in the same period last year to RMB 180 million, driven mainly by a RMB 100 million foreign-exchange loss. AIX Finance noted that net cash outflow from operating activities reached RMB 290 million in the first half, versus just RMB 110 million a year earlier, while receivables and notes receivable rose from RMB 960 million to RMB 1.11 billion. On embodied AI, the company disclosed only that "orders have achieved a breakthrough," without giving specific figures, shipment volumes, or revenue contribution; co-founder Chen Xi said in July that the Gino 1 is scheduled to enter small-batch production and delivery in the third quarter of this year. The stock closed at HK$10.2 on August 28, below its HK$16.8 IPO price and down nearly 70% from its January high.Source: Jingbao Net source, Chaoqi Net source

OpenAI product VP draws a falsifiable line for "affordable L5"

Peter Welinder wrote a lengthy post on X arguing that affordable L5 autonomous driving is still 5 to 10 years away, on the grounds that it requires AGI, which doesn't yet exist: first there needs to be multimodal, multiplexed AGI, then real-time inference capability, and finally all of that has to fit onto a power-efficient, inexpensive chip that can go into a car ordinary people can afford. Rather than stopping at the assertion, he laid out concrete thresholds: the vehicle should cost under $30,000, or a robotaxi's cost per passenger-mile should be under $0.65; there should be no geofencing, no remote operators, and no reliance on GPS or centimeter-level HD maps; the system needs to handle weather, faded lane markings, construction, and police hand signals, and cope with real-time negotiation at four-way stops in cities where lane markings haven't been repainted in a decade; and it needs to be able to recover from problems on its own, with no human takeover required. Musk's only reply was a clown emoji — no quote, no argument, no data.Source: OfficeChai source

XPeng's third Turing chip lights up in production vehicles, completing its three-chip architecture

Following the second-generation VLA's integration of 30-second memory and 6-second prediction into the model (previously reported), XPeng announced on August 27 that the third of its proprietary Turing-series chips has now lit up in production vehicles, with the three chips handling perception, decision-making, and cross-domain fusion respectively. Since mass production of the Turing chip began in Q3 2025, cumulative shipments have exceeded 200,000 units; the company says it will switch its entire lineup to the proprietary chip starting in Q2 2026, targeting annual shipments approaching one million units. The "super-agent" running on this hardware is a new version formed by merging the second-generation VLA and VLM, with an on-device model parameter count 3.5 times larger than the previous generation. End-to-end response speed is up 300%, enabled by streaming inference, in which the model receives input, reasons, and outputs a driving trajectory simultaneously rather than waiting for full processing to complete.⚠️ vendor claimSource: GlobalChinaEV source

Sanctuary AI starts selling its "robot brain" separately, with the humanoid body taking a back seat

This summer, on the production line of a global Tier 1 automotive parts supplier, the company plugged flexible harness connectors into moving targets with a reported success rate above 99.5% and a cycle time of 2.54 seconds. The footage shows no humanoid robot — just two suspended hands grasping wire, rotating, and inserting connectors. New CEO Daniel Friedmann explained how the figure was derived: testing ran for 40 minutes across 313 insertion attempts. The company's homepage now leads with "physical AI for industrial automation," and while the Phoenix Gen 8 humanoid robot still appears on the site, it now ranks below robotic arms, dexterous hands, and grippers. Rather than waiting for humanoid hardware to reach large-scale commercial deployment, the company is putting its models into existing industrial workflows and robots first — a hardware-agnostic approach that lines up well with Meta's real-world need for robots to plug in network cables and restart servers in data centers. The 40-minute, 313-attempt figure reflects only the scale of a capability demonstration, not proof of annualized production-line reliability.Source: Forbes source

Thirty LG executives fly to Santa Clara as physical AI partnership moves into technical talks

About 30 executives and business leaders flew to Nvidia's Santa Clara headquarters for a two-day session, holding multiple rounds of technical meetings on humanoid and logistics robot development while assessing commercialization feasibility. Under the plan announced on June 8, LG Electronics will use Nvidia's Isaac, GR00T, and Cosmos platforms to build out the full pipeline from data collection and simulation training to model deployment; on the manufacturing side, LG's production-line data will be connected with Nvidia's computing power and digital twin capabilities. The delegation included LG CNS president Hyun Shin-kyun, LG Electronics CTO Kim Byoung-hoon, and LG Innotek CTO John Min. News of the trip sent LG Electronics and LG CNS shares up as much as 14% intraday, with LG Corp also rising about 7% at one point. Yonhap reported that specifics of the first projects are expected to be announced only in a few months.Source: TradingKey source

Asus creates a Physical AI business group, renaming its AIoT unit on September 1

Asus's AIoT business group is being restructured and renamed the "Physical AI Solutions Business Group" (PAS BG) starting September 1, co-led by Jonney Shih (Zhang Quande) and Samson Hu (Xu Youjia), and organized into three divisions: AI Solutions, Autonomous Platforms, and Humanoid Robotics. Co-CEOs S.Y. Hsu and Vivian Hu wrote in an internal memo that as AI moves toward real-world systems, the company aims to stay ahead in the next wave of innovation. Asus estimates the total physical AI market will reach $1.1 trillion by 2035.⚠️ vendor claimSource: Taiwan Economic Daily source

First-half humanoid robot sales ranking: UBTECH, AgiBot, Unitree lead the pack

This ranking comes from the media account "Humanoid Robot Scenario Application Alliance," covering full humanoid units — bipedal, wheeled, and other mobile configurations, from full-size to compact — with figures estimated from a combination of public company data, historical production and sales trends, and market information. UBTECH led with 16,123 units sold in the first half (previously reported), followed by AgiBot in second place, which rolled its 15,000th unit off the line on June 28; Unitree came in third, saying in early June that cumulative production of its bipedal humanoid had reached about 11,000 units as of May. These top three companies are moving from the thousand-unit threshold toward tens of thousands, while most other companies remain at the small-batch validation and project-delivery stage.⚠️ compiled estimateSource: HK01 source

Hardware · Supply Chain

· Leaderdrive (Chinese harmonic reducer maker): plans to issue H-shares for a Hong Kong listing (previously reported); first-half 2026 revenue reached RMB 349 million, up 38.64% year-on-year, with net profit attributable to shareholders of RMB 70.1092 million, up 31.25%. According to QY Research, the company held about a 12% share of the global harmonic reducer market in 2025, ranking second worldwide; its share of the China market was 27.5%, ranking first. source

· Gaoteng Robotics: its Mecanum-wheel humanoid embodied chassis has already begun partnerships with several embodied-AI companies, with the company disclosing at the Shenzhen AGIC expo that shipments have reached 200 units⚠️ self-reported at trade show. source

FutureX · Physical AI Daily — Issue 104 (08/30)

Shawn — Sat, 29 Aug 2026 14:50:49 +0000

Today's Highlights

· Perceptron open-sources 36-billion-parameter embodied foundation model Isaac 0.5, single-demonstration learning outpaces π0.5

· Meta lets robots into AI data centers to swap cables, one employee estimates it could handle 80% of the work

· Caocao Mobility (Chinese ride-hailing platform) signs Abu Dhabi's K2, set to deploy its first overseas Robotaxi fleet this year

· NSF commits $30 million for a 39-person, six-university study on how humans and robots adapt to each other

· Xuanchuang Robotics (Chinese industrial robot startup) wins an order of over 100 units from PetroChina, delivering 20 per month, alongside a closed Series A1

· Robot at Xiangyang Science and Technology Museum injures girl by kicking her, operator reaches settlement with the family

Research Progress

Let the VLM point at a spot on the image, leave the geometry to SLAM · autonomy

At each step, the robot receives four surround-view images; rather than having the vision-language model compute 3D coordinates, it only needs to pick one of the images and mark a single 2D pixel point. The depth map and camera intrinsics then convert that point into a 3D position, which is handed off to the underlying SLAM system. Using this approach, the success rate on R2R-CE was pushed up to 66.2%. A Zhejiang University team argues the problem is that VLMs, from pretraining onward, only ever see 2D images and text — forcing them to compute 3D coordinates instead produces spatial hallucinations. In an ablation experiment they reverted to conventional metric waypoint output, and success rates on R2R-CE and RxR-CE dropped by 24.8 and 25.2 percentage points respectively, confirming this hypothesis. The approach also saves considerable overhead: each trajectory needs an average of only 9 interaction steps, taking 16.58 seconds per task, compared with 37.47 seconds for StreamVLN and 41.46 seconds for DualVLN. Chain-of-thought reasoning isn't run continuously either — it only triggers at key points like intersections or doorways, accounting for about 30% of the trajectory.

Hongyan Feng et al. (Zhejiang University ACES Lab × Zhejiang Humanoid Robot Innovation Center) · arXiv 2608.17512 source · Analysis: TechWalker source

Robotic fish builds a light field while questioning its own location · perception

Once underwater gliders submerge, they lose GPS signal and must rely on dead reckoning to estimate their position — a method whose error grows the longer it runs. A team at Michigan State University's approach is to treat this uncertainty itself as useful information: a multi-fidelity Gaussian process scores each measurement, and the less confident the robot is about its current position, the lower the weight assigned to that data point. They applied this approach to reconstructing a 3D light field in a large indoor water tank, bringing the final mapping error down to less than a quarter of that from conventional methods.

Demetris Coleman et al. (Michigan State University · JHU APL) · Autonomous Robots · Analysis: Bioengineer.org source

Prosthetic e-skin measures pressure and temperature simultaneously, modules snap together like Lego · perception

The fabrication process starts by scanning the prosthetic's shape, then laying out the sensor array to match that geometry, and finally using 3D printing combined with laser cutting to produce thin-film modules. The modules snap together via interlocking joints, with no adhesive needed. A Washington State University team says the system's sensing resolution reaches 10 times that of existing commercial glove sensors, and maintains high coverage density even on curved freeform surfaces. First author Hongyi Shen said this advance "makes the production of medical-grade e-skin accessible," laying the groundwork for a fully bionic skin that can both sense and deliver tactile stimulation.

Hongyi Shen et al. (Washington State University) · Cell Reports Physical Science · Analysis: Futurity source

Other papers today: ITRM uses trajectories "imagined" by a world model to give reinforcement learning agents early warning of environmental anomalies, without needing to inspect the policy internals — achieving an average AUROC of 0.853 across navigation, robotic control, manipulation, and autonomous driving tasks, with a false-positive rate of only 0.279 at a 95% true-positive rate.

Open Source · Tools · Benchmarks

· Isaac 0.5: Perceptron has open-sourced a 36-billion-parameter embodied foundation model, with weights on Hugging Face and fine-tuning code and technical documentation on GitHub. Training used over 3 trillion multimodal tokens, including 1 million hours of general video and 100,000 hours of robot experience, covering more than 35 robot embodiments. The company says the model integrates video understanding, embodied reasoning, and action generation into a single system; it scores 97.2% on the LIBERO benchmark, roughly on par with GR00T N1.7's 97.0% and π0.5's 96.9%. The real gap shows up in learning new tasks: given just one demonstration, error improves by 7.0–10.5x, versus only 2.3–3.1x for π0.5. The company was founded in November 2024 by Armen Aghajanyan and Akshat Shrivastava, both formerly of FAIR. source

· Benyuan Zhishu's "Dual Billion Plan": At the Big Data Expo, the company announced a five-year plan to build 1 billion expert trajectory data points plus 1 billion hours of embodied manipulation data, while also releasing four platforms: OriginBench for evaluation, OriginWise for expert data production, OriginFlow for agent operation, and OriginPhysics for embodied data collection. Founder Li Niyong said customers are no longer asking "how much data do you have" but rather "what kind of data will keep the model improving." Note that these figures are the company's stated targets, not its current scale. source

Funding and Deals

Xuanchuang Robotics ｜ Series A1 ｜ Tens of Millions of Yuan · industrial ⚠️ Company account

The investors are Qianhai Fangzhou, Guangyang Co., and West Lake Venture Capital; the funds will go toward data pipelines, training and deployment of VLA and JEPA world models, and stocking standardized products. Founded in December 2022, the company's core team comes from Harbin Institute of Technology and specializes in explosion-proof inspection for hazardous oil and gas/chemical settings, following an "inspection plus operation" approach. Its configurations include wheeled, rail-hung, tracked, and dual wheel-legged designs — notably, no humanoid form. Founder and CEO Fu Zhe disclosed that PetroChina's procurement approach has shifted from POC-project purchasing last year to bulk purchasing this year; an order of over 100 units signed with PetroChina's Xinjiang Oilfield division has already completed its first batch delivery, with subsequent deliveries proceeding at a pace of 20 units per month, and current orders on hand already exceed 100 million yuan. The first standardized final-assembly line came online on July 30, built jointly with Guangyang Co. Asked why the company doesn't build humanoid robots, Fu explained that chemical plant customers prioritize reliability far above form factor, and that a wheeled chassis with dual arms is "more reliable, more economical, and easier to deploy."Source: 36Kr source

Green Harmonic Drive (688017.SH) ｜ Plans H-Share Listing in Hong Kong ｜ Amount Undisclosed · hardware

The company announced on the evening of August 26 that its board has approved a proposal to issue H shares and list on the main board of the Hong Kong Stock Exchange, pending approval from shareholders, the China Securities Regulatory Commission, and the Hong Kong Stock Exchange. First-half revenue reached 349 million yuan, up 38.64% year-on-year; net profit attributable to shareholders was 70.11 million yuan, up 31.25%; non-GAAP net profit was 58.17 million yuan, up 36.92%. The company attributed this growth to an increased share of the industrial robot market and significant expansion of its embodied AI robotics business. Harmonic reducers and metal components contributed 288 million yuan, or 82.67% of revenue, with a gross margin of 32.79%; electromechanical products contributed 54.74 million yuan. The company was the earliest in China to achieve industrial-scale mass production of harmonic reducers, and holds roughly 15% global market share, ranking second worldwide and first in China.Source: Gasgoo source

EXL (EXLS.O) Acquires iMerit ｜ $310 Million ｜ Closed August 3 · adjacent

The deal consideration is split into two parts: $170 million upfront, plus up to $140 million in milestone-based earnout payments over the next two years. iMerit provides data services for AI model training, evaluation, and reinforcement learning, covering text, image, video, audio, and LiDAR data, serving robotics, autonomous driving, and other autonomous systems customers. Founder and CEO Radha Ramaswami Basu will become Executive Vice President of EXL and join its executive committee.Source: The Robot Report source

Commercialization and Deployment

Kinisi's KR1 installs crankshaft bearing shells at a UK auto plant, trained on just one day of human demonstration · embodied ⚠️ Company account

The bearing shells must be pressed into the crankshaft housing bore, with the press-fit force varying between 2,000 and 8,000 newtons depending on engine specification; if the fit is off, a camera alone can't detect it — only a change in resistance reveals the problem. This wheeled humanoid robot now handles three operations: picking and placing crankshaft components, fitting the bearing shells into their corresponding grooves, and completing precision-aligned press-fitting. It stands 162 cm tall and weighs about 100 kg; the customer is an unnamed global automaker, with the pilot running at Kinisi's headquarters in Bristol. CEO Brennand Pierce told UK tech outlet IZON that the automaker has identified over 15,000 similarly automatable operations across its global plants, with this single customer alone representing an addressable opportunity of over $2 billion; South Korean industry outlet TheGuru cited this report on August 27. Kinisi was acquired by BearRobotics (LG-owned robotics company) just ten weeks ago. Most previously publicized humanoid robot demonstrations have focused on warehouse logistics tasks, whereas precision engine component assembly has long served as a litmus test that automotive engineers use to judge whether embodied AI truly delivers manufacturing value.Source: Tech Times source

Caocao Mobility signs Abu Dhabi's K2, set to deploy first overseas Robotaxi fleet this year · autonomy ⚠️ Planned, unconfirmed

K2 is an AI technology company owned by the Abu Dhabi government; under this partnership, K2 will provide the testing environment, operational infrastructure, and local resources, while Caocao Mobility will supply intelligent customized vehicles, technical support, and training. The two sides plan to begin local testing and validation in 2026. This builds on Caocao's earlier cooperation with the Abu Dhabi Investment Office. Back in China, driverless testing without a safety operator began in Hangzhou's Binjiang district on July 27, using a station-triggered dispatch model with a local fleet of about 100 vehicles in Hangzhou, backed by Caocao Mobility's RAS remote safety service platform, which handles continuous monitoring, emergency call response, and remote takeover; back in April, it had already become the first company approved for driverless road testing in Hangzhou. China's first ground-up-designed Robotaxi model, the Eva Cab, has also been unveiled, with mass production set for 2027. Under the goals laid out in its RoboX strategy, the company aims to deploy a cumulative 100,000 Robotaxis and 100,000 Robovans by 2030.Source: scanx.trade source source

Einride to put 500 Tesla Semis on the road, rolling out across five states starting in September · autonomy

Texas, California, New Jersey, Illinois, and Georgia will be the first five states to gain access, with the trucks running on Einride's own fleet intelligence platform; customers include shippers at the scale of Amazon. The full rollout will proceed in phases over the next two years. Tesla Semi Program Director Dan Priestley said electric heavy-duty trucks lower per-mile costs through better fuel efficiency, reduced maintenance, and higher uptime, and that the economics will improve further once deployment scales up. Tesla has set the Semi's production launch event for September 24 at its Nevada factory, though large-scale deliveries of the vehicle have already been promised for several years.Source: CDLLife source

Tesla Robotaxi launches in Miami with no safety monitor on board · autonomy

Service currently covers only a small area of West Miami, with the map Tesla posted on X quite tightly bounded. This mirrors how the service launched in Houston and Dallas in April, both starting from very limited coverage areas, while Austin only expanded to citywide coverage last month.Source: Mashable source

Industry Developments

Meta puts robots to work in AI data centers, testing cable swaps and server reboots · industrial ⚠️ Single-source account

The robots being tested come from three manufacturers: San Francisco-based Watney Robotics, Quebec-based Kinova, and Zurich-based ABB, handling tasks including moving racks, inspecting equipment, swapping network cables, and rebooting servers. According to WIRED, an unnamed employee estimated that a robot capable of reliably completing cable-swap tasks could replace up to 80% of the work for certain roles — though this is that employee's personal estimate, not an official Meta forecast, and the robots currently still work more slowly than humans. Obstacles, battery life, visual inspection, and dense cabling remain sticking points, and the process still requires human oversight throughout. In response, a Meta spokesperson said the US is in the midst of its largest infrastructure boom since World War II, facing a severe shortage of skilled workers, adding: "We need more workers, not fewer."Source: Decrypt, citing WIRED source

NSF commits $30 million to establish center on human-robot co-adaptation, led by UT Austin · embodied

The five-year project brings together 39 researchers across six universities, with a research focus on long-term coexistence: how both people and robots change after living alongside one another for months or even years. Center director Joydeep Biswas, associate professor of computer science at UT Austin, said the next real leap "isn't just about getting robots to perform more tasks, but about getting them to understand the people around them — their needs, preferences, values, and constraints." The research relies on a real-world deployment network called HERO, with sites set up in homes, dormitories, cafes, public museums, rehabilitation hospitals, and senior/assisted-living residences at UT Austin, MIT, Yale, Indiana University Bloomington, the University of Utah, and Tufts University, using long-term deployment to observe what happens once the novelty wears off. Industry partners involved include Amazon, Apptronik, Diligent Robotics, Google DeepMind, Hello Robot, MassRobotics, NVIDIA, and Robust AI.Source: Weidu.net source

Over 70 embodied AI training grounds completed and operating nationwide in China, with 46 more under construction or planned · adjacent

This figure is as of the end of June 2026. Guizhou's first embodied AI "data factory" recently began operations, featuring 16 fully realistic simulation scenarios across 6 major categories — research, dining, industrial, retail, home, and office settings — following a complete pipeline of "scene capture → equipment calibration → cleaning and annotation." Zhang Jianwei, a foreign academician of the Chinese Academy of Engineering, said at the Big Data Expo that large-scale deployment of embodied AI is at a critical "window period," requiring data companies and robotics teams to jointly conduct training and validation in order to bring down innovation costs. At the same exhibition, China Southern Power Grid showcased a robotic dog rated to carry 200 kg of payload and climb 60-degree slopes.Source: China News Service source

Robot at Xiangyang Science and Technology Museum injures girl by kicking her, operator reaches settlement with family · humanoid

The robot was in a calibration/rehearsal phase at the time of the incident. Hefei Anda Chuangzhan Technology Co., the museum's operations service provider, said the compensation details would not be disclosed publicly; the robot involved was a Unitree unit. According to the museum, Unitree robot performances are still being held regularly.Source: Southern Metropolis Daily source

JD.com acquires land in three cities within two months to build RoboBases, targeting 80 within five years · adjacent ⚠️ Planned, unconfirmed

On August 26, JD.com won a bid for an industrial plot in Wuxi's High-Tech Zone, with a planned floor area of about 112,000 square meters; earlier, a Guangzhou Huangpu project broke ground on July 11 with a planned area of 190,000 square meters, expected to be completed by the end of 2028; a plot in Shanghai's Pudong Zhangjiang area spans 105,000 square meters. Each RoboBase is meant to combine display and delivery, maintenance, R&D and design, pilot-line assembly, data collection, and manufacturing functions all in one; iiMedia Research founder Zhang Yi likened it to a robotics "4S dealership," and said outright that JD.com's bet isn't primarily on hardware R&D itself. Zheng Xiaodan, head of JD Retail's embodied AI business, said the robotics industry boom isn't a moment you can wait for — it has to be actively built.Source: National Business Daily source

Galbot (Chinese embodied AI startup) unveils its first bipedal humanoid ET1, targeting a price of about $95,000 · humanoid ⚠️ Company account

After three years of building wheeled robots that pick items in pharmacies and convenience stores for handoff to delivery riders, ET1 marks the first time this Beijing-based company has put a robot on two legs. The unit stands 173 cm tall and weighs about 65 kg, with 48 degrees of freedom overall, including 24 across both hands; its compute platform is NVIDIA's Jetson Thor, and it carries an IP54 protection rating. It has already demonstrated fully autonomous play against human athletes in a sporting event (previously reported); the company's new figures put reaction time at 0.1 seconds, achieving roughly a 90% forehand contact rate against incoming balls traveling over 50 km/h. TechEBlog notes that most of these specifications come from aggregated data sheets rather than an official full spec release, and that the company itself still regards ET1 as a prototype.Source: TechEBlog source

ICE plans to spend up to $2 million on robot dogs · adjacent

The stated purpose is use in immigration enforcement operations. The figure comes from a government funding document reviewed by NBC News.Source: NBC News source

Hardware · Supply Chain

· Harmonic reducers: industry analysis estimates these account for about 30%–35% of a humanoid robot's total cost, concentrated mainly in lighter-load joints such as the wrist, elbow, and fingers source

· Jinlu Electronics (301282): PCBs used in integrated joint modules for humanoid robots have moved from trial production to mass production; the company stated on an investor platform that this business still accounts for a small share of revenue and has not yet had a material impact on results source

· Laifu Harmonic Drive: robot component revenue grew sharply, and adjusted losses narrowed source

FutureX · Physical AI Daily — Issue 103 (08/29)

Shawn — Fri, 28 Aug 2026 14:49:27 +0000

Today's Highlights

· Anthropic releases hardware MCP, letting AI agents directly control lab instruments and production line equipment

· UBTECH (Chinese humanoid maker) sold 16,123 units of humanoids in H1, full-size model revenue up 1,445%

· a16z raises $1.1 billion "American Dynamism"-style fund focused on AI's physical infrastructure

· Unitree's G1 EDU exposed with two root vulnerabilities, one starting from Bluetooth

· China's National Development and Reform Commission cools robotics enthusiasm: prevent blind bandwagon-jumping and herd behavior

· Riemann-1.0 hits 85% long-horizon real-robot success rate, 15 points above the strongest open-source baseline

Paper Progress

Riemann-1.0: compressing robot policy and world simulator into a single causal autoregressive model · world-model

Long-horizon real-robot manipulation success rate of 85.0%, process success rate 94.4%, 15 points above the strongest open-source baseline. Existing world action models mostly rely on joint generation, video-first approaches, or decoupled modeling; Riemann-1.0 instead places multi-view observations, robot state, and proprioceptive actions into a single causal sequence, so the same set of weights can both run policy online and serve as a multi-embodiment visual simulator. Training used over 200,000 hours of interaction data, drawn from a mix of first-person human video, handheld gripper demonstrations, and heterogeneous robot trajectories, following a progressive embodiment pretraining approach. In simulation: 94.3% on RoboTwin2.0, 99.0% on LIBERO, and 62.6% on the long-horizon composite benchmark RoboCasa-365.

Haofeng Sun et al. · arXiv 2608.27033 source

WALL-SS: "next-scale" autoregression extends world model usability to minute-scale horizons · world-model

World models have mostly been stuck at clip-level future prediction, drifting once switched to continuous interaction. WALL-SS writes embodied trajectories as a causal sequence interleaving observations and actions, generating each future frame coarse-to-fine across scales, paired with scale-compressed long-range memory: recent interactions keep fine resolution while distant observations and actions are compressed, keeping memory bounded. The authors report coherent streaming rollouts of over one minute, with on-policy alignment used to suppress action drift and long-horizon inconsistency.

Maeve Zhang et al. · arXiv 2608.26239 source

CLAP: cross-embodiment video world model doubles directly as a zero-shot physics simulator · world-model

Action-conditioned video models are usually locked to a single embodiment, leaving internet-scale human video unusable. CLAP reconciles different action spaces using three representations — end-effector pose, language instructions, and latent actions — first learning physical priors on unlabeled video, then grounding to end-effector action space for zero-shot deployment. On DROID it approaches or exceeds single-embodiment SOTA, covering forms including Bridge, dual-arm YAM, and the G1 humanoid, with code and models fully open-sourced.

Kechen Liu, Ola Shorinwa · arXiv 2608.27406 source

SOLO: a humanoid completes a continuous 1.5 km outdoor route using only a single chest-mounted depth camera · locomotion

Perception-based humanoids tend to fail over long distances, for two compounding reasons: dense terrain reconstruction smooths away critical detail, and point-wise imitation lacks temporal credit assignment. SOLO uses Fourier-encoded grid queries to extract evidence from depth and proprioceptive tokens, cutting elevation-map L1 error to a quarter to a third of the baseline; it also writes teacher-student next-state divergence into the PPO reward. In stress-test terrain, average pass rate was 97.5%, 96% on stepping stones, versus 0–3% for the dense-reconstruction version.

Pihai Sun et al. · arXiv 2608.26583 source

PredVLA: a 0.68M-parameter predictive-coding policy outperforms a same-scale transformer by sevenfold · vla

Large models dominate manipulation leaderboards, but it's unclear how much scale language-conditioned control actually needs. PredVLA has only 0.68M trainable parameters, uses no robot-data pretraining at all, and relies on hierarchical generative recurrent dynamics to predict visual features and proprioception, with observations only affecting the hidden state through online inference of prediction error. It averages 86.9% across three short-horizon LIBERO suites, 75.4% including the long-horizon suite. Under identical frontend, demonstrations, and evaluation protocol, it beats a parameter-matched Transformer policy by 3.7x and an LSTM by 7.4x.

Hiroki Sawada, Shunichi Kasahara (Sony CSL) · arXiv 2608.26673 source

Instruct-to-Act: a VLM issues low-frequency instructions while a world-model controller takes over at high frequency · vla

Instruction-tuned VLMs are good at translating instructions into high-level plans but poor at low-latency action sequences in unfamiliar environments; world-model controllers are the opposite — fast but lacking open-ended task guidance. This work relabels controller-policy rollout segments with synthetic instructions, optimizing a behavior-cloning objective jointly with the original reward and world-modeling objectives. Three of the seven embodied environments are multi-agent scenarios where VLMs coordinate via language and a trained controller acts as the executor; swapping in a different pretrained VLM planner requires no fine-tuning.

Zineng Tang et al. · arXiv 2608.26788 source

Learning five juggling patterns with three balls on a real robot in five minutes · manipulation

Models remain useful even when the sim-to-real gap is large. This online learning framework learns local models from accumulated experience while retaining a global prior model to extrapolate in experience-sparse regions, then constructs a mutually reachable set that guarantees transitions between catches never push the arm past joint or actuator limits. A dual-arm robot with multi-fingered hands and onboard vision learned five three-ball patterns — cascade, tennis, half-shower, shower, and box — in under 5 minutes of real-world interaction. Authors include Christopher G. Atkeson and Alfred A. Rizzi.

Taeyoon Lee et al. · arXiv 2608.26800 source

TrapVLA: backdoors that don't just make a robot fail, but dictate how it fails · vla

Prior backdoor attacks on VLAs treated any task failure as a success; this paper requires the attacker to control the failure mode itself — for instance, making a grasp miss by a specified distance — which raises the bar for detection. The authors build a data engine for synthesizing target trajectories along with an automated failure-fidelity evaluation, construct two benchmarks — Trap-LIBERO and Trap-RoboTwin — covering four failure-mode categories, and design a method that explicitly learns trigger-induced action residuals. In both simulation and real-robot experiments, performance on clean data is largely preserved.

Jun-Hui Liu et al. · arXiv 2608.26578 source

Other papers today: FlashVLA uses streaming action decoding to cut inference latency for flow-matching VLAs (arXiv 2608.27384 source); FLARE adds failure detection and autonomous recovery to VLAs (arXiv 2608.26645 source); TemporalFlow-VLA physicalizes execution history to resolve ambiguity between visually similar states in multi-stage manipulation (arXiv 2608.26821 source); Memory Anchors reduces catastrophic forgetting in continually-learning robot policies (arXiv 2608.26545 source); 4DSynth procedurally synthesizes interactive, time-varying embodied simulation scenes (arXiv 2608.26947 source); SpatialCrafter builds walkable worlds from a single image using a generative 3D agent (arXiv 2608.27073 source); RTNav achieves real-time zero-shot object navigation (arXiv 2608.26496 source); MeshPriorDiT hierarchically models action-conditioned cloth dynamics (arXiv 2608.26766 source); Generative Semantic Scene Completion recasts outdoor LiDAR semantic completion as discrete diffusion (arXiv 2608.26737 source).

Open Source · Tools · Benchmarks

· Hugging Face Microduck: a $399 (~¥2,681) open-source bipedal duck-shaped robot, 25cm tall, 770g, 15 degrees of freedom, with a Rockchip RK3566 chip and a monocular camera, plus an open-source SDK, virtual training environment, and reinforcement-learning scripts. Thomas Wolf told Bloomberg that pre-orders were selling roughly one unit every 4 seconds, with sales of about $500,000 and a target of over 20,000 units; the first batch ships before Christmas, manufactured by Seeed Studio (Chinese hardware maker) in Shenzhen source

· Amap ABot-Recon: Alibaba's Amap released a real-time 3D reconstruction solution, claiming to reconstruct scenes with tens of thousands of frames' worth of detail from just 12 input frames source

Funding & Deals

a16z "American Dynamism" fund ｜ New fund ｜ $1.1 billion · adjacent

The money targets the entire physical foundation of AI — chips, memory, interconnects, data centers, robotics — with a16z describing it as "opening the throttle to accelerate AI's physical buildout." A firm known for backing software-scale returns is putting a new fund behind hardware, and the fund's announcement named several gaps: faster and more efficient systems, cheaper and higher-bandwidth memory hierarchies, more scalable interconnects between nodes, and low-power edge devices that let AI explore and interact with the physical world.Source: TechCrunch source

Sharpa ｜ Cumulative funding ｜ over ¥4.5 billion · embodied ⚠️ vendor claim

Industrial investors including Alibaba, Meituan, Tencent, JD.com, and Transsion, alongside institutional investors such as Sequoia China, Qiming Venture Partners, Meituan Longzhu, and Photosynthesis Capital, make up the shareholder list of this dexterous-manipulation company founded in 2024. The figure comes from the company's own disclosure, without a breakdown by round or timing. Sharpa was founded by three co-founders of Hesai Technology (Chinese lidar maker) striking out on their own; co-founder Li Yifan is simultaneously Hesai's CEO. Its flagship offerings are the Sharpa Wave dexterous hand and a companion full-stack robot "brain."Source: Sina Finance source

Mech-Mind ｜ Hong Kong IPO offering ｜ Priced at HK$95.30–101.70 ｜ up to HK$2.706 billion raised · industrial

After the book opened on August 24, the cornerstone investor lineup has emerged: Baillie Gifford leading a group of nine, alongside Taikang Life, Jane Street, BYD's investment arm Golden Link, Invus, Ghisallo, Ernst & Young, NGS Super Fund, and E Fund. Approximately 23.14 million shares are being offered. The company builds the 3D-vision "eyes," AI-decision "brain," and execution "hands" for robots; per a CIC report, it held roughly 22.1% of the global market for AI+3D-vision-guided non-dedicated intelligent robot components in 2025, corresponding to about ¥389 million in revenue — nearly 2.6 times the second-place player's 8.5% share.Source: ACN Newswire source

InstAdapt ｜ Seed round ｜ tens of millions of dollars · world-model

Co-invested by Xiechuang Zhihui, Yunhui Capital, Pudong Science and Technology Investment, and Wuyue Angel, with Xinpai Capital serving as financing advisor. The company was incubated from a ShanghaiTech University team; founder Shi Ye is an assistant professor at ShanghaiTech's School of Information Science and Technology and head of YesAI Lab. Its focus is embodied world-model infrastructure — letting robots switch environments, tasks, or even embodiments without collecting data and training from scratch. Its technical approach was invited to be presented at NVIDIA GTC 2026.Source: PEdaily source

Psi Robot ｜ New round ｜ over $100 million (~¥672 million) · embodied

Following yesterday's disclosure of the round, today the investor list was made public: industrial investors including Tuopu Group, Chery Holding's Ruicheng Fund, and Lens Technology (Chinese electronics component maker), alongside institutional investors 37 Interactive Entertainment, Wuhu Municipal Investment Holding Group, and Fosun Chuangfu, with existing investor Huajin Capital adding an over-allotment. The company's prior angel and Pre-A rounds together totaled ¥2 billion. This funding is primarily aimed at world models: the company's self-developed Psi-R2 manipulation policy model handles long-horizon task decomposition and continuous action planning, while its Psi-W0 action-conditioned world model handles trajectory prediction and correction.Source: cnbeta source

Lingzhang Technology ｜ Angel round ｜ tens of millions of yuan · embodied

The funds will go toward industrializing its dexterous-hand technology.Source: PEdaily source

Xingji Miyuan ｜ Seed round ｜ tens of millions of yuan · embodied

Led by Vertex Capital China, with Woan Robotics and Xuyuan Capital participating. The company positions itself as a supplier of embodied-intelligence systems for all-domain scenarios.Source: Jiwei.com source

Commercialization & Deployment

UBTECH sells 16,123 humanoid units in H1, full-size model revenue up 14x · humanoid

Total revenue of ¥1.27 billion, up 104.2% year-on-year. Total humanoid robot sales of 16,123 units, up 268.3% year-on-year; of which full-size embodied-intelligence humanoid robot revenue was ¥590 million, up 1,445.0% year-on-year. Overall gross profit was ¥570 million, gross margin 44.7%, up 9.7 percentage points year-on-year; the expense ratio fell to 54.4%. The company is still losing money, but the losses are narrowing: adjusted EBITDA was -¥170 million, a 45.9% reduction in loss year-on-year. R&D spending exceeded ¥300 million, with 1,103 R&D staff. The sales figure includes models beyond just full-size units; the company reported the two figures separately.Source: Sina Finance source

DQ's Shanghai Wujiang Road store hands Blizzard-making over to a robot, 55 steps with no human operation · embodied ⚠️ vendor claim

This DQ store in Shanghai, opening August 29, is a collaboration between Sharpa and CFB Group; without altering equipment, ingredients, or process, the robot uses cups, spoons, and blenders designed for human use to complete all 55 steps of making a Blizzard in one continuous run, with Sharpa claiming tactile sensing is directly involved in 98% of them. The above process and figures come entirely from the company's press release, with no independent record yet of sustained operation; the store's long-term reliability will only be known once it has actually been running for a while.Source: Sina Finance source

ICE plans to spend $1–2 million on Boston Dynamics robot dogs · adjacent

A U.S. Immigration and Customs Enforcement notice within the Department of Homeland Security states it will procure Spot robots and accessories for "inspection, situational awareness, and hazard assessment," citing improved officer safety. Spot starts at around $75,000, implying at least several units under the stated budget range. The same day's procurement records also show ICE ordering 6,000 pairs of electroshock gloves for $16.7 million. A Boston Dynamics spokesperson told 404 Media that the company sells to many government agencies and public-safety organizations for uses including hazardous-gas detection, unexploded-ordnance inspection, suspicious-package investigation, and search and rescue, and stressed that any attempt at weaponization is explicitly prohibited under its terms. The New York Police Department previously canceled a similar contract following public backlash.Source: 404 Media source

Thousands of L4 logistics robots from Chengdu's Yangshi Technology now running regular delivery routes on European streets · autonomy ⚠️ vendor claim

These robots are already handling commercial orders for food delivery, supermarket services, and last-mile logistics transfer on public roads in several overseas cities. Each unit has over 60 liters of cargo capacity, a 20 kg payload, supports quick battery swaps, and offers over 12 hours of range per charge; it has 6-wheel drive with independent suspension, can clear obstacles up to 14 cm, and reaches a top speed of 10 km/h. The company plans to deploy 20,000 units nationwide in a single overseas country by the end of 2027 — this is a target, not confirmed orders. Yunda Technology disclosed on an investor-relations platform that Yangshi Technology is its affiliate, in which it holds a 36.7% stake.Source: Guandian.cn source

Japan's Muse delivery robots enter a supermarket in Queens, New York · industrial

Armo One went into service this month at Foodcellar Market, carrying up to 100 kg of goods; staff use an app to summon a loaded robot from the back storage room to the shelves, and after swapping modules it can also scan for out-of-stock items and check price tags. It's billed on a subscription basis, as low as $1 to $2 per hour per unit. Muse's U.S. business head, Nobuaki Momoi, said "about 20% of restocking time goes to moving things around, like pushing carts." Store owner Metin Mangut has been unable to fully staff the store since the pandemic, with some positions going unfilled despite postings; he said, "we believe technology should support people, not replace them." Most U.S. automation investment is concentrated in less-visible places like e-commerce warehouses; retail robots that customers actually encounter remain rare. MarketsandMarkets projects the U.S. physical AI market growing from roughly $300 million in 2026 to about $2.7 billion by 2031.Source: Nikkei Asia source

Industry Developments

Anthropic releases MHS hardware standard, letting AI agents directly drive lab instruments and production-line equipment · adjacent

The Model Hardware Standard is aimed at any device with a programmable interface, covering both scientific instruments and advanced manufacturing equipment; Anthropic compares it to USB-C — a unified way for devices to exchange information. It is not tied to Claude and is model-agnostic; the company says the goal is to shrink the time it takes enterprises to connect and integrate hardware. It's currently in research preview, open only to a handful of organizations in science, robotics, and manufacturing, with plans to open-source it later so that hardware makers in any industry can adopt it. Elizabeth Kelly, who leads beneficial deployments at Anthropic, told CNBC, "We built this for science, to show what AI's potential looks like, but there's enormous value here for enterprise and industry too." The company open-sourced MCP in 2024, which has since become the de facto standard for agents connecting to data sources; it is also building a silicon team for its own chips, recently hiring Caitlin Kalinowski, who previously worked on hardware at OpenAI, Meta, and Apple.Source: CNBC source

China's NDRC: robotics industry must fit local conditions, prevent blind bandwagon-jumping and herd behavior · embodied

Li Chao, deputy director of the National Development and Reform Commission's Policy Research Office, said at an August regular press briefing that the robotics industry touches on numerous frontier technologies including AI, advanced manufacturing, and new materials, and that each region should find its positioning based on local resource endowments and industrial strengths. The next focus areas are embodied-intelligence training grounds and application pilot bases, letting robots iterate on technology in real-world settings. On data, the goal is to build a high-quality real-robot data collection system to address training-data scarcity; on models, the plan is to support companies pursuing multiple technical paths and encourage innovation across vision, language, action, and world models to accelerate technical convergence. On applications, the goal is to expand the library of embodied training scenarios, encourage companies to open up real-world scenarios, and support reliability and safety testing, extending into manufacturing, healthcare, consumer, service, and public-safety sectors. This stance stands in noticeable contrast with the wave of robotics industry plans that various localities have issued in recent weeks.Source: Cailian Press source

Unitree's G1 EDU exposed with two independent root-RCE chains, one starting from Bluetooth · humanoid

Security researcher Olivier Laflamme disclosed CVE-2026-76639 and CVE-2026-76640 on August 27. The first exploits a path-traversal condition in chat_go to reach bashrunner, gaining root on the Locomotion PC. The second starts from BLE proximity: the initial write path doesn't require Bluetooth pairing, and Unitree's cloud service at the time only verified that the requester was a valid Unitree account, not that the account owned the target robot — allowing an attacker to retrieve key material for another G1 EDU, use it to establish an authenticated BLE state, and enter the Wi-Fi provisioning code, where a buffer overflow also leads to root. Unitree patched the cloud-side ownership check in July 2026, which Laflamme says broke his original verification workflow. The exact patched firmware version could not be determined, leaving G1 EDU users without a clear upgrade target; whether the two vulnerabilities affect other models is also unconfirmed. Propagation testing was limited to two robots in the same room.Source: The Hacker News source

Rokae turns profitable in H1, embodied-intelligence business now a third of revenue · industrial

Revenue of ¥415 million, up 136% year-on-year, adjusted net profit of ¥18.4 million, with operating cash flow turning positive. Total robot sales exceeded 13,000 units, up 225% year-on-year. Broken down: industrial robot revenue up 69%, flexible collaborative robot revenue up 72%, and embodied-intelligence robot revenue jumping to ¥139 million — 34% of total revenue — with sales exceeding 5,600 units. Management attributed this to multiple embodied-intelligence companies adopting or switching to its humanoid robotic arms for R&D and product iteration. Gross margin was 29.8%, up 7.5 percentage points year-on-year; R&D spending was about ¥63.9 million, up 146.6% year-on-year.Source: Sina Finance source

Hanoi approves 24-month autonomous-driving pilot, driverless buses, robotaxis, and delivery robots all hitting the road together · autonomy

A decision by the Hanoi People's Committee on August 25 designated two zones. Within the Hoa Lac High-Tech Park, an approximately 11.5 km route will run robotaxis and delivery robots for on-demand passenger transport and delivery; within the Hoa Lac campus, an approximately 2.8 km route will run cleaning and delivery robots, alongside driverless buses ferrying students. Operations run daily from 9:00 to 16:30, with a speed cap of 40 km/h, and are suspended during heavy rain, thunderstorms, flooding, strong winds, or dense fog. Driverless buses and robotaxis must be staffed with a safety operator able to take over at any time, with passenger limits of 8 and 4 respectively. Commercial fares are prohibited during the pilot. The city will temporarily waive registration, licensing, inspection, and operating-permit requirements within the pilot zones, replacing them with alternative safety measures.Source: VnExpress International source

Waymo launches $29.99/month membership, turning robotaxis into a subscription · autonomy

Waymo Premier is invitation-only; its core benefit is 10% back in Waymo Cash per ride, with higher cashback during busy periods. Other perks include priority pickup, priority access when new cities launch, and up to 5 free cancellations per month. Waymo says this was requested by users. At $29.99/month, the price implies riders need to use it frequently enough to break even.Source: Mashable source

Hardware · Supply Chain

· LinkTech (Chinese harmonic reducer maker): harmonic reducer capacity has reached 1 million units, targeting 1.5 million by year-end, with over 60% market share in China; order visibility extends to Q3 2027, with humanoid-related orders making up over 40% of the total; its weight-reduced version for humanoids is 20–45% lighter than the standard version. On August 26 its board approved a proposal to issue H-shares and list in Hong Kong source

· Yeoyo Technology: joint orders on hand for 2026 have surpassed 1.3 million units, with full-year deliveries projected at 1.3 million — versus just 95,000 units shipped in all of 2025; its completed automated joint production line has an automation rate of 85%, with a 60-second cycle time per harmonic joint source

· Frameless torque motors: Leadshine Technology's on-hand orders for humanoid frameless torque motors exceeded 1 million units as of H1, with a production line targeting an annual capacity of 3 million units also being built out; Lingqiao Drive & Control's on-hand orders for frameless motors passed 1 million units in H1, versus just 120,000 for all of 2025 — its general manager said, "the biggest pressure right now is delivery" source

· Six-axis force sensors: Kunwei Technology has annual capacity of 60,000 units, targeting 120,000 by year-end, already supplying UBTECH, AgiBot, Galbot, and Xiaomi at scale; Xinjingcheng's 6.5mm- and 9.5mm-diameter micro six-axis force sensors are already deployed at scale in dexterous-hand projects, with Q1 shipments exceeding 100,000 units and a full-year target of over 500,000 source

· Gap between planned capacity and actual delivery: the year-end capacity figures and full-year delivery targets above are all public plans, not yet realized; LinkTech's own capacity-expansion project via a follow-on share offering is only 8.39% complete, with completion pushed back two years source

· Planetary gear reducers: DigiTimes reports that humanoid robot demand is driving up shipments of planetary gear reducers source

· Nanshan Zhishang: its self-developed high-strength transmission tendon rope achieved small-batch orders in H1 source

FutureX · Physical AI Daily — Issue 102 (08/28)

Shawn — Thu, 27 Aug 2026 15:01:24 +0000

Today's Highlights

· SoftBank in talks to acquire controlling stake in 1X, valuation $6 billion, five years after selling Boston Dynamics

· XPeng's second-generation VLA builds time into the model, remembering the past 30 seconds and projecting 6 seconds ahead

· XPeng Robotaxi secures Guangzhou remote-testing qualification, entering driverless front-seat road testing

· Waymo shares ten lessons from 200 million miles: L2-to-L4 upgrading is a "false summit"

· South Korea to invest 2.3 trillion won in full-stack humanoid robotics by 2030

· Tiangong Ultra clocks 8.64 seconds in the 100m final; AgiBot (Chinese humanoid robotics company) tops medal table with 46

· Yikong Zhijia (Chinese autonomous mining-truck company)'s driverless mining trucks double to 3,100 units, yet revenue falls 17.3%

Research Papers

One policy runs arms, humanoids and human hands: Xiaomi team's camera-centric action space · vla

VLA models aim to consume all robot data, but are stuck on incompatible embodiments, camera setups and underlying action spaces — the standard industry fix is a separate adaptation branch per dataset. Xiaomi's Embodied Intelligence team and the University of Macau took a different approach: instead of treating robot commands as the shared target, they redefine actions using anchor motion visible to the camera, so that arms, humanoids and human hands all become different carriers of the same action pattern, with a geometric-conditioning translator combined with the target embodiment's kinematics to recover executable control. Trained on 4.03K hours of robot and simulation data plus 2.34K hours of human demonstrations, a single checkpoint with no task-specific fine-tuning scored 98.3% on LIBERO, 82.0% zero-shot on LIBERO-Plus, and 89.2% on RoboTwin's hard scenarios.

Shaoqing Xu et al. (Xiaomi Embodied Intelligence Team · University of Macau) · arXiv 2608.26058 source

Zero-WAM: robots learn unseen tasks from a single human video used as a prompt · vla

Large language models can perform new tasks purely from context, but robots can't — the bottleneck is "task specification" itself. This work argues that the most natural specification for manipulation tasks isn't language but human video, since visual cues directly describe how the task should unfold. To address the scarcity of paired data, the team built an automatic pipeline converting robot trajectories into semantically matched human videos, assembling HumanGen (74.2K pairs covering 8.6K tasks), and used in-context future-clip prediction to discourage the model from taking shortcuts. Across seven unseen tasks in RoboTwin 2.0, average success rate reached 47.0%, 29.5 percentage points above the strongest video-action baseline.

Jiaming Zhou et al. · arXiv 2608.26103 source

R³: getting robots to reason through hard problems in natural language before acting · vla

Language reasoning has bought foundation models extra test-time compute, but whether that mechanism transfers to manipulation tasks was previously unclear. R³ is a two-stage post-training recipe: first mid-train an off-the-shelf VLM on expert reasoning trajectories to lock in a reasoning style, then apply single-step rubric RL using offline action data. Unlike prior work that treats structured trajectories as auxiliary supervision, it trains free-form language reasoning whose output directly guides a low-level policy. Across two testbeds — Language Table and a simulated dual-arm bagging task — both exploration and generalization on unseen tasks beat instruction-only imitation-learning baselines.

Lehong Wu et al. · arXiv 2608.26053 source

StreamPI: adding temporal memory to single-frame VLA without adding a single parameter · vla

Hugging Face trending score of 15 and rising. The strongest current VLAs, such as pi0.5, still operate frame-by-frame, with no memory of what just happened. StreamPI treats each (visual observation, language instruction) pair as an atomic time unit: bidirectional attention within a unit handles cross-modal fusion, while causal attention across units preserves autoregressive streaming inference, with the language instruction serving throughout as a semantic anchor. Training randomly samples frame intervals (e.g., every 3rd frame), yielding robustness to the frame-timing jitter of asynchronous real-robot deployment. On real-robot memory and fine-perception tasks plus LIBERO, it outperforms pi0.5 while directly inheriting single-frame pretrained weights.

Zhe Liu et al. · arXiv 2608.26067 source

TacForcing: touch can still change its mind mid-execution of an action chunk · manipulation

Contact state can change within a single action chunk, but chunk-based VLA tactile conditioning is frozen at the moment before execution — effectively working with stale information. Prior fixes bolted on a separate high-frequency reactive controller, complicating both architecture and training. TacForcing instead replaces the standard action expert with a streaming version, letting action generation track tactile observations arriving continuously during execution, and uses Execution-Aware Tactile Attention to apply tactile conditioning only to the steps nearest execution. Across six UniVTAC simulated tasks and three contact-rich real-robot tasks, average success rates reached 65% and 69% respectively.

Jianbo Zhou et al. · arXiv 2608.25798 source

MA-VLA: breaking cooperation into atomic instructions so multi-arm coordination can recombine · manipulation

Most VLAs treat language as a single global instruction with no mechanism to specify which arm does what, so they collapse on cooperation patterns absent from training. MA-VLA decomposes cooperative behavior into mid-level atomic prompts assigned to each arm, and introduces Arm Shuffle during training — permuting each arm's observations, state and atomic prompt — to force the model to execute instructions role-agnostically. The team also built a test set whose cooperation patterns are entirely absent from training, on which previous SOTA VLAs largely fail. Code, model and data are public; Hugging Face trending score of 5.

Zaibin Zhang et al. · arXiv 2608.25864 source

Twelve physical AI benchmarks turn out highly redundant, two pairs can simply be merged · benchmark

Model evaluation sets rarely overlap across papers, the model×benchmark matrix is sparse, and no one had measured whether benchmarks actually test the same thing. The authors selected 51 models × 12 benchmarks by reporting density from a registry of 51 benchmarks and 152 models, mixing model-card scores, benchmark-paper scores and their own reproductions, and found significant redundancy: merging two pairs of substitutable benchmarks into single columns shifted the equal-weighted average ranking of 22 of the 51 models by more than three positions. Greedy selection based on score dispersion plus variance unexplained by the already-selected set retained 78.5% of the utility of all 12 benchmarks using just four.

Zaruhi Navasardyan, Hrant Davtyan · arXiv 2608.25940 source

Super Odometry 2.0: holds pose through smoke, sandstorms and low light · perception

Odometry systems tend to collapse in smoke, sandstorms, snow or low light — precisely the daily conditions of search-and-rescue and mining. This version elevates the IMU to equal standing with cameras and LiDAR in sensor fusion, letting it take over when exteroceptive sensors fail, paired with a hierarchical adaptive mechanism for varying degrees of degradation. This is the most-cited paper in today's batch, with 12 citations.

Shibo Zhao et al. · arXiv 2608.25427 source

SkyDrive: watching traffic from drones overhead to teach autonomous driving new cities · autonomy

Planners trained via imitation learning often see large zero-shot performance drops in a new city, as traffic rules, road networks and driving habits all change — and re-collecting local data means deploying an entire vehicle sensor suite. SkyDrive instead uses drones for traffic-monitoring supervision: vehicle-mounted collection only records the ego vehicle and its surroundings, whereas a single aerial pass can observe large numbers of road users simultaneously, turning every vehicle into a data source. Drawing 650,000 driving samples from 137 hours of aerial footage and building a benchmark from it, experiments show that just 30 minutes of monitoring per location noticeably narrows the cross-city gap for most models. Data and code will be released.

Weijiang Xiong et al. · arXiv 2608.25142 source

Other papers today: GaussianDream++ adds metric 3D supervision to VLA via 3D Gaussian reconstruction and future prediction (arXiv 2608.25659 source); ConfAL-WM selects samples by confidence for post-training action-conditioned world models, improving efficiency on RoboTwin2.0 (arXiv 2608.25572 source); V-Link fixes VLA's action expert lacking access to visual representations (arXiv 2608.25308 source); VISTA infers spatial contact attention from vision, beating 3D Diffusion Policy on cap-twisting and calligraphy tasks (arXiv 2608.25872 source); LAC gives a humanoid upper body simultaneous linear and angular compliance (arXiv 2608.25405 source); RA-VLA does test-time adaptation via retrieval augmentation (arXiv 2608.25585 source); PRISM uses a GPU physics simulator as an online world model for dual-arm MPC (arXiv 2608.25666 source); 4DStreamCtrl unifies camera motion, object trajectories and depth into a 3D point-track representation (arXiv 2608.25479 source); CRESSim-Neo is a batched GPU simulation engine for surgical robots (arXiv 2608.25192 source); VBVR-Pro turns native visual reasoning into a verifiable closed-loop testbed (arXiv 2608.26105 source).

Open Source · Tools · Benchmarks

· Microduck: Hugging Face's second robot, priced at $399 and shipping this year, developed by French company Pollen Robotics (acquired last year). About 25cm tall and 770g, with LiDAR, camera, microphone and dual NFC readers, programmable in Python or JavaScript, and can be taught new skills via reinforcement learning and simulation. Its predecessor, Reachy Mini, has sold roughly 10,000 units to date. source

· DEEPRobotics (Chinese robotics company)'s DM0.5 tops RoboDojo: following this month's report that the top model on this real-robot leaderboard — launched by the University of Hong Kong and nearly 20 other institutions — scored just 12.8%, DM0.5 took first place with a composite score of 24.90 and average success rate of 19.34%, with its Memory-dimension score of 47.74 opening the gap; LIBERO composite success rate was 99.0%. The model is open-sourced. source

· NVIDIA COMPASS: a cross-embodiment navigation policy training framework that uses an AI agent to generate training curricula, aiming to help robots learn to navigate new environments faster. source

· South Korea opens 1.56 trillion tokens of AI training data: government-led dataset construction, made publicly available. source

Funding & Deals

1X Technologies (Norway) | Controlling-stake talks | Valuation ~$6 billion · humanoid ⚠️ Reported, unconfirmed

The Information, citing people familiar with the matter, reports that SoftBank is in talks to acquire a majority stake in this OpenAI-backed humanoid robotics company; neither side has confirmed it publicly. The timing is worth noting: just over a month ago, Hyundai announced it would buy SoftBank's remaining roughly 10% stake in Boston Dynamics, ending SoftBank's ties to Spot and Atlas entirely. When SoftBank sold an 80% controlling stake in Boston Dynamics to Hyundai in 2021, the whole company was valued at roughly $1.1 billion. Five years later, Masayoshi Son is now looking at a younger, earlier-stage-commercialization company priced at more than five times that. SoftBank's previous in-house robotics venture ended with Pepper, discontinued in 2020 after a cumulative production run of about 27,000 units.Sources: The Information source, 36Kr source

Ren Shaoqing's new company | NIO strategic investment | Unicorn-level valuation · embodied ⚠️ Single-party account

On August 24, NIO CEO William Li announced at an autonomous-driving all-hands meeting that Ren Shaoqing, NIO's head of intelligent driving, has founded an independent company for physical AI foundation models and embodied intelligence, with NIO making a strategic investment and entering a collaboration; Ren will continue heading NIO's intelligent-driving business. Ren Shaoqing is one of the four core authors of ResNet and a co-author of Faster R-CNN; in 2016 he co-founded Momenta, and in 2020 he joined NIO to build its autonomous-driving team from scratch. This "stay-in-role while founding a separate company" arrangement has no established precedent, and questions remain about how his time will be split under the dual role, and where the boundaries lie for vehicle data and intellectual property. The same day, XPeng's robotics unit announced a first round of over $900 million in funding (previously reported).Source: Robot Insight source

Psi Robot (灵初智能) | New round | Over $100 million · embodied

Investors include Tuopu Group, Chery Holding's Ruicheng Fund, Lens Technology, 37 Interactive Entertainment, Wuhu Municipal Investment Holding Group and Fosun Chuangfu, with existing investor Zhuhai Science & Technology Industry Group adding an above-allocation top-up. The company builds embodied "brain" models, with a proprietary dual-system architecture combining the Psi-R2 manipulation policy model and Psi-W0 action-conditioned world model — the former handling long-horizon task decomposition and continuous action planning, the latter forecasting and correcting action trajectories. Funds will go toward two areas: scaling up human-data collection, including building data-collection facilities across multiple locations; and pushing deployment in logistics and advanced manufacturing. Its previously reported automated optical-module production line, built jointly with YOFC, uses robots for sub-millimeter-precision insertion quality checks and vacuum sealing.Source: Gasgoo source

Yikong Zhijia (7687.HK) | First half-year report since listing | Revenue RMB 549 million · autonomy

The number of operating driverless mining trucks rose from over 1,600 a year earlier to over 3,100 (previously reported), yet revenue fell 17.3% year-on-year. The reason lies in the gross margin: the product mix shifted toward the lighter-asset "customer-provided fleet" model, whose share of revenue rose from 62.7% to 65.6%, pushing half-year gross profit up from RMB 11 million to RMB 104 million, a gross margin of 18.9%. The company posted a loss of RMB 189 million for the period, down from RMB 236 million a year earlier. It serves 38 mines, with 4 sites running fleets of over 200 trucks and the largest single-mine fleet at 566 units; over 1,000 additional trucks are on order and in the delivery pipeline. Cash and cash equivalents on hand stood at RMB 273 million as of June 30.Source: LeiDi source

Wonik Robotics (South Korea) | National Growth Fund investment | 350 billion won · industrial

Following this month's report of a planned raise of over 300 billion won, South Korea's National Growth Fund has formally finalized its investment, with the funds going toward building a new Jeonbuk AX Center. Also approved in the same batch was CJ 4DPLEX, with the two equity investments totaling roughly 570 billion won.Source: Seoul Economic Daily source

Motion (Brussels, Belgium) | Pre-seed | €1.7 million · humanoid

Offers "humanoids-as-a-service," supplying humanoid robots to European manufacturers on a service basis; the company says it's targeting Europe's robot-adoption gap rather than building its own hardware. Roughly $2 million.Source: EU-Startups source

Xijian Technology | Nearly RMB 500 million raised over the past year | Neuromorphic chips · hardware

Incubated out of Tsinghua University, builds neuromorphic chips.Source: ZDX (智东西) source

Commercialization & Deployment

LG Innotek mass-produces robotaxi camera modules for Zoox · hardware

5-megapixel modules, with different fields of view depending on their mounting position on the vehicle, together forming a 360-degree detection ring around the vehicle body; they're waterproof and must maintain performance in severe weather and sharp temperature swings. This marks LG Innotek's first mass-produced camera module built specifically for robotaxis; the two companies have collaborated since Zoox's early development, and the modules will be installed on vehicles over the coming years. Zoox completed a robotaxi factory in Hayward, California last year and is already running driverless ride-hailing in Las Vegas. LG Innotek is also working on LiDAR and radar, aiming to grow its autonomous-driving perception solutions business to 2 trillion won (about $1.45 billion) by 2030 and expand its customer base to automakers and robotaxi companies in North America, Europe and Japan.Sources: Businesskorea source, The Korea Times source

XPeng Robotaxi enters driverless front-seat road testing, completes 2,000 rides in internal trial within two months · autonomy

The Robotaxi, equipped with the second-generation VLA, has obtained remote-testing qualification for intelligent connected vehicles from the city of Guangzhou, allowing testing without a safety operator in the driver's seat on relevant grade-1, grade-2 and grade-3 test roads in Guangzhou. Since launching an internal employee trial in June, it has completed over 2,000 rides in two months, covering complex urban roads. The test fleet, based on the XPeng G9 platform, has been running in Guangzhou for five months since May. XPeng is pursuing a unified technical foundation, scaling the same base model up to L4 Robotaxi and distilling it down for production vehicles with lower on-board compute.Source: Kuaikeji source

Didi unveils Robotaxi R2; Beijing-Guangzhou demonstration zones now open to the public · autonomy

Co-developed with GAC Aion, the vehicle was formally unveiled at the 18th International Transport Exhibition. It carries 33 sensors, including 10 LiDAR units, cameras, 4D millimeter-wave radar, an infrared camera and sound sensors; GPU compute exceeds 2,000 TOPS, and Didi says its tri-domain-fusion central computing "brain" is the industry's first mass-produced solution of its kind. The vehicle is built on GAC's GEP 3.0 pure-electric platform, with electronic steering control and redundant braking. Portions of Huangpu district in Guangzhou have begun round-the-clock trial operation, with users hailing rides through the "autonomous driving" entry point in the Didi app; once it obtains a Beijing road-testing license, the R2 will join Beijing's "Huiju Port" operations network. The demonstration zones remain limited in scope, still far from full-area rollout.Source: WTOutlook (网通社) source

LimX Dynamics (Chinese humanoid startup)'s bipedal humanoid shipments pass 3,000 units · humanoid ⚠️ Company-reported

The company released operating figures on the closing day of the humanoid robot games: cumulative global shipments of its bipedal humanoids have surpassed 3,000 units, with over 2,000 shipped so far in 2026 and over 1,200 delivered in the first half of the year, putting it, by its own claim, among the top three globally. Shipments and deliveries are two different measures — the 3,000 and 2,000 figures are shipments, while the 1,200 figure is deliveries.Source: Guandian.cn source

Hyundai begins delivering IONIQ 5 autonomous vehicles to Waymo in Q4 · autonomy

The vehicles are produced at Hyundai's plant near Savannah, Georgia, then shipped to a Magna-operated retrofit facility to install the Waymo Driver, with a rooftop sensor suite including LiDAR, radar, cameras and microphones, plus additional sensors on the front and rear bumpers and front fenders. Waymo's previous-generation fleet is being replaced.Source: Sina Finance source

Turkey's Arvato deploys Lipbo (Chinese sorting-robot maker) 3D sorting robots, 3,700 packages per hour · industrial

Arvato, a Bertelsmann-owned logistics service provider, deployed the system in Turkey, with a processing capacity of 3,700 packages/hour.Source: Weidu.net source

Industry Developments

XPeng's second-generation VLA builds a time dimension into the model, rolling out in September · autonomy

It "remembers the world from the past 30 seconds" and projects scenarios 6 seconds into the future. At XPeng's Physical AI sharing day on the afternoon of August 27, the company named the capability behind its second-generation VLA Infini-VLA, built on a physical-world foundation model that for the first time incorporates a time dimension — officially described as AI's understanding of the world jumping from 3D space to 4D spacetime. Traditional inference is a serial loop of observe, compute, output, observe again; the new version switches to streaming autoregressive inference — observing, thinking and acting simultaneously — cutting end-to-end response time by 300%. The on-vehicle model's parameter count grew 3.5x, which the company says is more than 15 times that of mainstream industry VLA models, with multidimensional overall safety capability up 20x. In a live demo, the system remembered that a vehicle ahead was mid-U-turn, held back cautiously while waiting, and quickly moved off once the vehicle cleared. The new version rolls out in September to all Ultra and Ultra SE models, debuting first on the G9L; a distilled version, Turing VLA 2.0 Lite, will roll out the same month as a first batch to Max models with a single Turing chip. XPeng frames this version as reaching up to L4 while remaining compatible with more mass-production models, unifying L2 through L4 on a single technical foundation.Sources: Kuaikeji source, Guancha source

Waymo shares ten lessons from 200 million miles, aimed squarely at the L2-to-L4 upgrade path · autonomy

"Improving a driver-assistance system into full self-driving is a false summit. True L4 maturity can only come from safety built into the system by design, validated on closed courses, and hardened by the rigorous experience of driving with no one in the vehicle." Waymo's head of AI foundations, Srikanth Thirumalai, posted ten lessons on Tuesday, without once mentioning Tesla by name. The first lesson states that cameras alone cannot achieve safe autonomous driving at scale — Waymo uses cameras, LiDAR and radar as redundancy; the fourth targets pure end-to-end networks that map raw footage directly to steering commands: "you cannot build trust with a black box." Waymo currently provides over 500,000 paid fully driverless rides per week, targeting 1 million by year-end; Tesla confirmed last month that its unsupervised FSD had covered just 380,000 miles over a full year. The same day, XPeng announced it was unifying L2 through L4 on a single technical foundation — the two companies reaching opposite conclusions about the same technical path.Source: Electrek source

Humanoid robot games close: 100m in 8.64 seconds, AgiBot tops medal table with 46 · humanoid

Following this month's successive record-breaking times of 9.39 and 8.86 seconds, X-Humanoid's Tiangong Ultra brought the 100m down to 8.64 seconds in the closing-day final, 0.22 seconds faster than its repechage run, and nearly a second faster than Usain Bolt's human world record of 9.58 seconds. The fastest 100m time at last year's inaugural event was still over 12 seconds. Its closest rival in the final was Ubtech (Chinese humanoid robotics company)'s Lightning. Robots and humans compete under entirely different physiological, technical and rule conditions, so the times aren't directly comparable. AgiBot, competing for the first time, took 46 medals including 18 gold, topping both the gold-medal and overall-medal tables. X-Humanoid CEO Xiong Youjun said after the event that the "ChatGPT moment" for embodied intelligence is imminent.Sources: Interesting Engineering source, Robotics & Automation News source

Reuters feature: China's humanoid robots aren't yet smart enough to take your job · humanoid

Hardware leads the world, but intelligence can't yet support general-purpose work — that's the scene Reuters reporter Laurie Chen and colleagues observed at a training facility in Guangxi. The site has over 100 humanoid robots generating training data, moving slowly and clumsily; a novice trainer needs roughly 300 attempts to produce one usable motion, while a skilled trainer needs 50. UBTECH won an $18 million procurement contract from the Guangxi provincial government last October, supplying humanoid robots and related hardware for the site. Chinese humanoid shipments topped 40,000 units in the first half of the year, a 97% global share (previously reported) — but set against this usable-motion production rate, the lead in shipments and the gap in intelligence are two separate things.Source: Reuters source

AgiBot releases WALL-SS autoregressive world model · world-model ⚠️ Company-reported

Departing from diffusion models that generate video clip-by-clip, WALL-SS uses "next-scale autoregression" instead: sketching a coarse-scale outline first, then progressively refining image and physical detail. All three benchmark figures come from the company itself. Action-following score reached 0.29, which the company says is the only nonzero score among all models compared, with alignment between generated video and given trajectory commands rising from 0.251 to 0.539. Streaming rollout runs stably for 60 seconds without collapsing; by contrast, a model retaining only the most recent frames begins showing disappearing objects and warped tabletops after 20 to 30 seconds. The same policy showed a task-success-rate consistency of 0.93 between virtual and real-world settings, measured across six tasks and five training stages. The model has a built-in action-decoding head that reads the next action directly from a predicted frame, letting the same network serve as both simulator and planner.Source: Kuaikeji source

Teradyne sues JAKA's German subsidiary at the Unified Patent Court in Europe · industrial

The case was filed with the UPC's Copenhagen local division, involving software and hardware patents held by Teradyne Robotics A/S and its Universal Robots unit, targeting a range of collaborative robot models that JAKA sells in the EU. Teradyne Robotics president Jean-Pierre Hathout said, "we cannot and will not accept companies that illegally copy protected technology." This is the company's second IP lawsuit in Europe in 2026, following an earlier German suit over Universal Robots software copyright against another robotics company's subsidiary. Any ruling could apply across 17 of the 18 EU member states participating in the UPC, with the UK and Spain also named in this case.Source: Robotics & Automation News source

Hyundai considers selling Boston Dynamics robots through its car dealership network · humanoid ⚠️ Remarks at an event

Hyundai's CEO discussed the possibility of selling humanoid and quadruped robots through existing dealership channels, without giving a timeline or pricing. Also disclosed at the same event: Hyundai plans to deploy over 25,000 robots across its own factories by 2028.Sources: The Times of India source, CarsGuide source

Renesas opens Physical AI and robotics lab in Beijing · hardware

Focused on chip and system-level solution development for next-generation robotics.Source: Business Wire source

Hardware · Supply Chain

· AgiBot dexterous hands: the company says its critical-point series has shipped over 20,000 units cumulatively, holds the largest global market share, and has accumulated over 50,000 hours of real-world data; the OmniHand that won 7 golds at the games belongs to the same line. These figures are company-reported and have not been independently verified. source

· Zhongke Lingxi (Chinese robotics company) LG-603 industrial dexterous hand: global debut. The company says the prior-generation platform ran continuously for over a year at a Lili Motor factory in Lu'an, with 10,000 hours of zero-fault operation — roughly equivalent to 300,000 work cycles. The unit measures 203×82×40mm, weighs 510g, has a static load capacity of 25kg, a minimum grip diameter of 1-2mm, and a maximum overall grip force of 85N. A lighter variant, the LGX-601, weighs 400g with a 10kg static load capacity, and has opened a 30-day presale. The reference point is the 10,000-hour service-life threshold for dexterous-hand servo joints set by the Ministry of Industry and Information Technology's "Humanoid Robot and Embodied Intelligence Standards System (2026 Edition)": ≥10,000 hours. source

· Group standard for wheeled humanoid mobile chassis: jointly initiated by 14 companies, currently recruiting co-drafting participants. source

· Xiamen builds out dexterous-hand industry: a local industrial initiative, with products aimed at a range of fine-motor-skill scenarios. source

This Week's Watch

South Korea to invest 2.3 trillion won by 2030 in full-stack humanoid robotics · humanoid

South Korea's Ministry of Economy and Finance released this plan ahead of next year's budget announcement: 2.3 trillion won (about $1.66 billion) to be invested by 2030, with 600 billion won allocated for next year alone. The targets are specific: build a mass-production system for humanoid robots for 10 major industries; raise the localization rate of key components from the current 45% to 80%; procure 250 domestically made humanoid robots in 2027 for distribution to universities and state-funded research institutions, reaching a cumulative 1,080 humanoids by 2030, plus roughly 5,000 total including quadrupeds and other non-humanoid platforms. R&D focus is set on three components where South Korea's competitiveness lags — actuators, robotic hands and sensors — plus high-performance AI chips and batteries for humanoid robots. The government says it aims to complete full-stack humanoid technology, including a proprietary robot operating system, by 2031.Source: Korea JoongAng Daily source

US adds foreign-made advanced robots to FCC covered list · industrial

This month's previously reported domestic-component-ratio requirement now has its full rule text. The FCC Public Safety and Homeland Security Bureau's designation took effect on July 28, 2026, adding humanoids, quadrupeds, AMRs and other wheeled or tracked ground platforms as a category to the covered list under the Secure and Trusted Communications Networks Act. Being listed means the products cannot obtain FCC equipment certification — effectively barring import, marketing or sale in the United States — unless granted conditional approval by the Department of Defense, with an application deadline of January 1, 2028. To avoid this classification, a robot must be manufactured in the US and meet a domestic-component cost threshold, which is 65% through 2028 and rises to 75% in 2029. Models that already received FCC certification before July 28, 2026 are unaffected. ARC Advisory published an analysis this week on the practical impact for industrial robots.Sources: FCC official FAQ, Morgan Lewis, ARC Advisory source

Three market forecasts, three different bases: humanoid sales, robotaxi fleets, cost-parity point · adjacent ⚠️ Compiled from multiple sources

Robot Insight cites a Morgan Stanley forecast that Chinese humanoid robot sales will grow 133% year-on-year to 28,000 units in 2026, potentially reaching 2.6 million by 2035. The same article cites a Soochow Securities research note giving a cost threshold: once a robot's unit price falls below RMB 400,000, replacing human labor becomes economically viable. On the robotaxi side, Businesskorea cites Counterpoint Research data projecting a global robotaxi fleet of 3.6 million vehicles by 2035, with a service market worth $168 billion. The three figures come from three different organizations using three different methodologies and don't corroborate one another. Soochow Securities separately published an in-depth computing-industry report this week on the re-rating potential for industrial software and automation as world models shift from "generating the world" to "controlling the world."Sources: Robot Insight source, Businesskorea source, Faxian Baogao source

This week in supply chains: dexterous hands move from the arena to the production line, service-life becomes the new bar · hardware

The most concentrated supply-chain activity this week centered on end effectors. The robot games featured a dedicated dexterous-hand event, with a group of manufacturers including AgiBot, LimX Dynamics and Booster Robotics (Chinese humanoid startup) using it as a marketing showcase; at the WRC show, 373 companies and 311 new products debuted, with component makers unusually taking center stage — audience questions shifted from "what motions can it do" to "how long can it run reliably." The line that actually got drawn came from the Ministry of Industry and Information Technology's "Humanoid Robot and Embodied Intelligence Standards System (2026 Edition)," which sets a dexterous-hand servo-joint service-life requirement of ≥10,000 hours. Zhongke Lingxi's claim of 10,000 hours zero-fault operation over a full year at the Lu'an Lili Motor plant, and AgiBot's reported cumulative shipment of 20,000 units, both benchmark against this line — but both are self-reported figures with no third-party verification available yet. Pressure is coming from both directions: integrators developing end effectors in-house, cross-industry entrants flooding in, and capacity expanding too fast, squeezing component suppliers in between. South Korea's 2.3-trillion-won plan naming actuators, robotic hands and sensors as its three weak points also suggests that the difficulty of replacing these three components sits at roughly the same order of magnitude across countries. No material price or capacity changes elsewhere upstream to report this week.Sources: China.com source, NetEase source, Korea JoongAng Daily source

FutureX · Physical AI Daily — Issue 101 (08/27)

Shawn — Wed, 26 Aug 2026 15:03:06 +0000

Today's Highlights

· Waymo sets end of 2027 for opening robotaxis to the public in Munich, its first entry into the EU

· China's Ministry of Industry and Information Technology lists intelligent robotics among six emerging pillar industries, with embodied intelligence named a future industry

· Yikong Zhijia (Chinese autonomous mining truck company) now operates over 3,100 driverless mining trucks, double a year ago

· Figure launches crowdsourced data-collection app, covering 108 countries with 16 million videos uploaded so far

· NVIDIA Jetson Orin Nano 2 doubles inference performance while cutting power draw 40% at equal performance

· Generalist's valuation rises to $3 billion, with 8VC leading a $200 million round expansion

· TrAct uses visual trajectories as an interface, lifting real-robot success rate from 49% to 76%

Research Progress

E-skin splits into fast and slow paths: familiar touch gets a millisecond reflex, uncertain touch gets handed to a language model for slow thinking · perception

The skin itself decides whether touch should be handled by reflex or by reasoning. Familiar contact is classified in milliseconds via spike-based encoding, while uncertain contact is routed into a language-guided reasoning pathway, with confidence determining which path is taken. An accompanying commentary by Wenzhuo Wu of Purdue notes that this confidence-based routing design balances speed and reliability in unpredictable environments.

Sun et al. · Nature Sensors (published online August 26, 2026) · Commentary source

TrAct treats visual trajectories as an intermediate interface for control and prediction · world-model

Robot actions are tied to the robot's own body and don't align with changes in image space — a longstanding problem for conditioning signals in world models. TrAct has the VLA model output candidate actions together with corresponding visual trajectories; a world model then simulates the visual outcome of each trajectory, and a vision-language reward model picks the one that best matches the instruction for execution. Against the π0.5 baseline, LIBERO-INTEGRAL success rate rose from 27% to 55%, and real-robot Franka manipulation rose from 49% to 76%. Li Fei-Fei is among the authors.

Zhi Cao et al. · arXiv 2608.24101 source

Do world models really follow the action? Swap in non-expert trajectories and the cracks show · world-model

Using a world model as a learned simulator assumes it produces a faithful future for any valid action — an assumption that has never been tested. WorldEcho probes this using visual integrity and SE(3) trajectory alignment, finding that existing models handle expert actions reasonably well but, when fed non-expert trajectories, either ignore the instruction or generate visually implausible outcomes. A companion method, WorldSync, addresses this via distribution coverage, representation grounding, and intervention-effect alignment, yielding higher success rates for policy iteration on RoboTwin and real robots.

Sixiang Chen et al. · arXiv 2608.24885 source

NVIDIA builds a surgical world model that humans can control in real time · world-model

Generative surgical simulation has previously been unable to run in real time. Cosmos-H-Dreams distills a bidirectional teacher into a causal few-step student via Self Forcing, achieving streaming inference at roughly 160 FPS on a single RTX PRO 6000 Blackwell; it is control-agnostic, drivable directly from a browser keyboard, a Meta Quest, or CMR Surgical's Versius surgical console, and learned policies can also run in closed loop. The team calls it the first interactive surgical world model supporting real-time control by both humans and policies, and it has been released as an open simulation system.

Javier Gamazo Tejero et al. (NVIDIA) · arXiv 2608.24199 source

Moving "imagination" off the critical path: a 48-millisecond action chunk that's more accurate than the synchronous version · world-model

Generating video synchronously can't keep up with control frequency, but dropping imagination altogether hurts success rate. GlanceWAM splits the two jobs inside a single video DiT: an asynchronous proposer runs on a slow clock in the background, glancing at a single future frame seconds ahead, while the action head decodes purely in latent space in 48-millisecond chunks without being blocked. Across 24 RoboCasa kitchen tasks it reaches 72.2%, above the synchronous Cosmos Policy's 67.1% and the no-imagination co-trained baseline's 64.4%, and reaches 99.0% on LIBERO at 24x the speed of the synchronous baseline. Code has been released.

Linhan Wang et al. · arXiv 2608.23927 source

Inference latency can break standard RL entirely — this asynchronous scheme fixes it · vla

Larger VLA models are slower, and slowness changes the environment's effective dynamics, which breaks the Markov assumption that RL relies on. ARLI augments an asynchronous inference framework with two additions: it folds the already-committed action and a mid-inference observation into the state, restoring an approximate Markov structure. Across simulated and real-robot manipulation tasks, standard RL fails completely under latency, while ARLI not only fine-tunes successfully but matches or exceeds standard RL under an idealized latency-free setting.

Brian Zhu et al. (Siemens · UC Berkeley) · arXiv 2608.23831 source

VLA models assume every gripper is the same — 103,000 demonstrations disprove it · vla

Parallel-jaw grippers and suction cups require different strategies to achieve the same grasp goal, yet existing VLA datasets are overwhelmingly parallel-jaw. MiGA spans multiple robots covering five gripper types and collects 103,000 demonstrations, explicitly recording strategy divergence for the same goal; the accompanying GVLA uses a multi-gripper tokenizer plus adapter-based policy routing, outperforming baselines on zero-shot and few-shot adaptation to new objects and unseen tasks.

Hanyi Zhang et al. · arXiv 2608.24603 source

Vision-tactile fingertip durability improved two orders of magnitude — swap the cartridge when it wears out · perception

Commercial GelSight Mini and DIGIT sensors ruptured after just 24–30 seconds in an abrasion test, and after 25–35 minutes under repeated compressive loading. This fingertip covers the silicone layer with an unpigmented, textured TPU protective film paired with a replaceable sensing cartridge, taking about 2–3 hours to reach rupture under the same conditions; under repeated loading at 39.2 newtons and 45 cycles per minute, nine test samples all remained functional when testing was stopped, with seven lasting the full 5 days. Damage is gradual and causes minimal interference with tactile imaging at end of life. Edward Adelson is among the authors.

F. Richard Cottrell et al. · arXiv 2608.24242 source

Other papers today: CARO embeds a fixed-base Euler-Lagrange model into the RL loop, achieving zero-shot robust quadruped control without torque sensors (arXiv 2608.24217 source); a fiber-optic sensing glove uses multi-core shape-sensing fiber to reconstruct full-hand pose, with fingertip error of 7.2mm dropping to 4.9mm after factory calibration (arXiv 2608.24572 source); Concept-Guided Exploration builds sustainable, executable scene graphs and selects hazard-mitigating actions in real time, 6 citations (arXiv 2608.23650 source); a survey of first-person intelligent smart-glasses platforms proposes an L0–L5 grading scheme, HF↑8 (arXiv 2608.24877 source); NeoWorld-Pro reframes monocular reconstruction as procedural generation of interactive 3D scenes (arXiv 2608.24212 source); CAT proposes trajectory-level continuous action representations, decoupling action representation from control frequency (arXiv 2608.24111 source); PonderPounce directly reuses an MLLM's native causal context as robot memory (arXiv 2608.24115 source); GaussianWAM distills the geometry and semantics of 3D Gaussian fields into world-action models (arXiv 2608.24714 source); Latent Action as Intention replaces explicit future-observation generation with latent actions, cutting WAM latency (arXiv 2608.24882 source); SIREN-Bench generates and evaluates emergency-vehicle interaction scenarios by behavior (arXiv 2608.24094 source).

Open Source · Tools · Benchmarks

· KUAVO-VLA: LejuRobot (Chinese humanoid maker) open-sources an industrial-vertical VLA that adds a layer between a general foundation model and task-specific skills. Fine-tuned on LingBot-VLA 2.0 using over 600 hours of same-embodiment real-robot data, it achieves 48.27% overall success across 25 tasks (20 custom industrial tasks + 5 GM100 tasks), versus 16.27% for the base model; π0.5, GR00T N1.7, and LingBot-VLA 2.0 were compared when selecting the base model source

· Pipette: Liquid AI open-sources an on-device model benchmark suite under Apache 2.0, treating "model × quantization × runtime × device" as a single unit of measurement. The initial dataset covers over 1,000 configurations and 30+ models, running llama.cpp builds on macOS/iOS/Windows/Android, validated on an M5 Max MacBook Pro, iPhone 17 Pro, and Galaxy S26 Ultra, with methodology independently reviewed by Artificial Analysis source

Funding & Deals

Generalist (USA) | Series B extension | Nearly $200 million | $3 billion post-money · world-model ⚠️ unconfirmed report

Following yesterday's disclosure of $200 million in new funding, the investor and valuation were confirmed today: 8VC led the round, expanding the $400 million Series B led by Radical Ventures in June at a $2 billion valuation to $600 million total, with regulatory filings confirming this is an extension. The company was founded in 2024 by former Google DeepMind researchers Pete Florence and Andy Zeng, and former Boston Dynamics engineer Andrew Barry, building robot foundation models that can drive different robot bodies, pairing off-the-shelf robotic arms rather than developing its own humanoid. The amount and valuation come from two people familiar with the matter; the company has not confirmed them publicly. In the same space, Physical Intelligence is valued at about $11 billion and Skild AI at $14 billion. Source: Tech Funding News source

GenRobot (Jianzhi Robotics) | Series A | Amount undisclosed | Over $200 million raised to date · adjacent ⚠️ company account

Momenta led the round, with existing investors following on. This company, which builds embodiment-agnostic data and evaluation infrastructure, is only a year old and has already completed 7 rounds. It has developed the Gen DAS unobtrusive wearable data-collection device and the Gen Controller skill terminal, recording human operations via sound, magnetic tactile sensing, piezoresistive sensing, capacitive sensing, flexible fabric, and force feedback across multiple channels; its self-trained Data Foundation Model covers pretraining dimensions including chain-of-thought, 6D pose, depth, tactile sensing, hand tracking, and full-body mesh. The company says it already has embodied models pretrained entirely on its embodiment-agnostic data, demonstrating cross-scenario transfer in complex manipulation generalization tests. Source: CYzone source

Futuring Robot | New round | Over RMB 1 billion raised across three rounds in half a year · embodied ⚠️ company account

This round drew nearly 10 investors, including ByteDance, CATL Capital, Guofang Venture Capital, Deneng Capital, and Nice Group, with existing investors also increasing their stakes. Founded in 2022, the company builds general-purpose household robots for consumers and has launched two models, F1 and F2; in May it released its new-generation robot brain, Self-Evolving WAM, and deployed it to F2, which at WRC 2026 completed a seven-step meat-sauce pasta task end-to-end. The company says its robots are now in regular operation in over 500 ordinary households nationwide, with a core technical team of nearly 200 people. Source: Gasgoo source

MileBot | Strategic financing | Nearly RMB 100 million · adjacent

Funded by CITIC Securities, with Haoguan Capital acting as exclusive financial advisor. Founder Chen Gong holds a PhD in biomedical engineering from the National University of Singapore, and the company's approach uses self-developed compliant actuators for active rehabilitation training, rather than rigid exoskeletons that passively move the patient. 36Kr calls it the first company in China's medical exoskeleton sector to achieve a working profit model. The company cites clinical validation from the First Affiliated Hospital of Nanjing Medical University and Shenzhen Second People's Hospital: stroke patients in the active-training group showed significant improvement over baseline by week 2, versus week 4 for conventional treatment. Next steps focus on brain-computer interfaces and home elder-care scenarios. Source: 36Kr source

SYNAPATH AI | Angel round | Nearly RMB 100 million · adjacent ⚠️ company account

Builds human behavior data products spanning Ego, Exo, Tactile, and other data modalities, with a pipeline running from collection, cleaning, and motion/spatial information extraction through to quality inspection and delivery. The company says it has accumulated several hundred thousand hours and over 36 million data entries, spanning more than 260 scenario types, and has partnered with embodied-AI companies including AgiBot, Fourier Intelligence, and Robot Era, as well as major internet firms, on data and skill transfer. Source: Gasgoo source

Lingyu Intelligence | Pre-A round | Several hundred million RMB · adjacent

Jointly invested by Zhuzhou Industrial Investment, Feitu Venture Capital, Future Boundary Venture Capital, and Xinneng Venture Capital — the company's third round this year. Founded in early 2025, it develops the Lingyu TA robot platform in-house, paired with a real-robot data-collection pipeline and cloud-based operations platform; the funds will go toward expanding mass-production capacity for the platform and upgrading the data-collection pipeline. The company says it has already achieved large-scale commercial deployment in open retail settings. Source: PEdaily source

Commercialization & Deployment

Waymo sets end of 2027 for opening robotaxis to the public in Munich, its first entry into the EU · autonomy

In the coming weeks, Waymo will begin phased road testing in Munich, with human drivers mapping routes, training autonomous specialists, and adapting the software to local driving habits, ahead of a public launch planned for the end of next year. This marks Waymo's third international city after London and Tokyo, and its first entry into the EU. It will start with a small all-electric fleet of Jaguar I-Pace vehicles before expanding. Co-CEO Tekedra Mawakana said Munich is a world-class hub for mobility and engineering, that earning community trust ahead of launch is the top priority, and that the company will invest in local fleet operations and create high-skill jobs. Waymo currently operates in 11 U.S. cities, with 19 more in progress; as of March this year, it had logged nearly 221 million fully driverless miles. Last month, Uber also announced a partnership with Pony.ai and Croatia's Verne to launch service in Europe, with Munich also on its planned list. Source: CNBC source, Silicon Republic source

Yikong Zhijia's driverless mining truck fleet tops 3,100, double a year ago · autonomy

The strongest number in the interim results wasn't revenue. As of June 30, the company had over 3,100 driverless mining trucks in operation, up from over 1,600 a year earlier, with more than 1,000 additional trucks ordered and in delivery. By the company's own research, China's total operating fleet is about 6,000 trucks, giving it a share of over 53%. It serves 38 mines, up 12 from a year earlier, including 9 mine sites with fleets over 100 trucks and 4 over 200, with the largest single mine at 566 trucks. Overseas, in a partnership with Zijin Mining's Norton Gold Fields and Thiess in Australia, the first 6 mining trucks entered real production at the Kalgoorlie Havana open-pit gold mine in Western Australia. Financially, the picture diverges: revenue was RMB 548.8 million, down 17.3% year-on-year, while gross margin rose from 1.7% to 18.9%, and net loss narrowed 20.2% year-on-year to RMB 188.8 million; the share of revenue from the customer-provided-fleet model rose from 62.7% to 65.6%. Source: Landsea Green Life source

Camera towers take over receiving inspection at a cold-storage facility, logging 64,000 receipts in six months · industrial

At Lineage's cold-storage facility in Decatur, Alabama, receiving has been handed to Kargo's fixed camera towers. As pallets enter the dock, they are imaged from multiple angles, and computer vision extracts label text, barcodes, batch numbers, expiration dates, and dimensions, writing directly into the warehouse management system, with damage or mismatched paperwork flagged in real time with photo evidence. The facility processes over 500 pallets a day; a task that used to take several minutes per receipt now takes seconds, and over 64,000 receipts have been logged in the past six months. Lineage's VP and Chief Data Scientist Elliott Wolf said automation frees up staff for higher-value work rather than replacing them. The hardware is rated to withstand forklift impacts, plugs into standard 110–120V outlets, and doesn't require Wi-Fi. Automated outbound loading scans are also set to launch at Decatur. Source: Unite.AI source

Chengdu-made L4 delivery robots hit European streets · autonomy ⚠️ planned figures

Yangsi Technology (Chengdu-based robotics company), whose mass-produced L4 autonomous logistics robots number in the thousands, has these robots already running regularly on public roads in several overseas cities, handling food delivery, retail, and last-mile transfer orders. The units have over 60 liters of cargo capacity, a 20kg payload, support fast battery swaps, offer over 12 hours of range per charge, use 6-wheel drive with independent suspension, can clear 14cm obstacles, reach a top speed of 10km/h, and sense their environment via lidar, cameras, and ultrasonic radar. The company plans to deploy 20,000 units across an entire overseas country by the end of 2027. Source: Sohu source

XPeng's robotaxi completes over 2,000 internal test rides in Guangzhou · autonomy ⚠️ earnings-call account

He Xiaopeng said on the Q2 earnings call on August 24 that the production robotaxi equipped with second-generation VLA has completed over 2,000 internal test rides in Guangzhou, validating the full passenger-demonstration workflow, with the cloud-based remote takeover platform now complete; the target is to begin carrying passengers without a safety driver in 2027. The second-generation VLA has already completed localized testing in Germany, aiming for one model to handle both Chinese and European road conditions. Q2 total revenue was RMB 19.74 billion, up 8.0% year-on-year, with 103,295 vehicles delivered and R&D spending of RMB 2.91 billion, up 32.1% year-on-year. Source: Gasgoo source

Industry Developments

Intelligent robotics enters the "15th Five-Year Plan" as one of six emerging pillar industries; embodied intelligence named a future industry · adjacent

At a State Council Information Office briefing, Vice Minister of Industry and Information Technology Xin Guobin laid out the industrial structure for the next five years: accelerating the buildout of six emerging pillar industries — integrated circuits, aerospace, biomedicine, the low-altitude economy, new-type energy storage, and intelligent robotics — while advancing quantum technology, biomanufacturing, hydrogen energy and nuclear fusion, brain-computer interfaces, embodied intelligence, and 6G as future industries and new growth drivers. The AI section covered four areas: supply, application, ecosystem, and safety. On the supply side, it named support for R&D of high-end training chips and multimodal algorithms, tackling frontier technologies such as brain-inspired intelligence and world models, and advancing intelligent terminals such as humanoid robots and brain-computer interfaces. On the application side, it proposed a "one industry, one map; one scenario, one profile" approach to scenario cultivation, aiming to distill decades of manufacturing know-how into lightweight, low-cost, easily deployable industry solutions. Yao Jun, director of the Planning Department, gave quantified targets: by 2035, the goal is to cultivate roughly 100 demonstration parks and 1,000 demonstration enterprises. The new-energy-vehicle section noted plans to actively and steadily advance pilot programs for intelligent connected vehicle access and road use, and for vehicle-road-cloud integration. Source: Guangzhou Daily / Xinhuacheng source

First dedicated dexterous-hand competition shifts the evaluation bar from "can it run and jump" to "can it do the work" · embodied

Eight competition events covered fine manipulation tasks such as powder weighing, tweezer bean-picking, cable connection, and tool assembly, with an average of 51 teams registering per event and only 7 to 17 advancing to finals. AgiBot's subsidiary Lingjiedian (Critical Point) entered its mass-produced dexterous hand, OmniHand, for the first time and swept awards in all 8 events, taking 7 gold, 4 silver, and 3 bronze; all 14 medals came from the same unmodified production unit — a 510-gram, 16-degree-of-freedom hand. The competition scored teleoperation and full autonomy separately, weighting autonomous performance at 1.0 versus 0.5 for teleoperation, though most teams still chose teleoperation. Among Critical Point's 3 autonomous gold medals, the powder-weighing win used algorithms developed independently by a Wuhan University team on OmniHand hardware. Another team, Lingxin Qiaoshou (Dexterous Hand), was the only team in the competition to compete fully autonomously across all eight events, driving in 18 screws in 5 minutes during the power-tool assembly final, missing points on only one screw for slightly exceeding torque tolerance, scoring 17 out of 18 to lead second place by 8 points, using an 11-degree-of-freedom, 370-gram Linker Hand O6. In the scenario events, Galbot (Chinese robotics company) competed fully autonomously using its self-developed AstraBrain world-action model, sweeping gold in the household, dining, and retail categories, finishing the household event with 305 points and the fastest completion time. Critical Point CEO Qiao Tianjie drew a line in an interview: for tasks involving impact loads and multi-step planning, such as fixed nail-driving or unpacking, they still had to fall back to teleoperation. Source: Sohu source, source, PR Newswire source

Figure crowdsources data collection worldwide, app now live in 108 countries · adjacent ⚠️ company account

On August 25 U.S. time, Figure launched Index simultaneously on iOS and Android, paying users worldwide to submit real first-person videos of manipulation tasks. It now covers 108 countries and regions, with over 264,000 downloads, more than 44,000 weekly active contributors, and over 16 million videos uploaded. The company says it will invest over $1 billion in data and compute over the next 12 months. Figure had previously relied on third-party data vendors, but found that scale, diversity, and quality couldn't keep up with what training its Helix model required.Source: 36Kr source

Jetson Orin Nano 2 doubles inference performance while cutting power draw 40% at equal performance · hardware

The module delivers 78 TOPS of compute, 8GB of memory, and an 8-core Arm CPU, keeping the same module form factor as the previous-generation Orin Nano Super. NVIDIA says the gains come from an improved Tensor Core and higher memory bandwidth; in 15-watt mode it matches the Super's performance while using 40% less power. On the software side it runs Jetson AI Lab's agent skills, supporting Cosmos, Nemotron, and edge-tuned models such as Gemma 4 and Qwen 3, with NVIDIA saying over 3 million developers have used its robotics stack. Deepu Talla, VP of Robotics and Edge AI, said, "today's small and medium frontier models have reached the accuracy of last year's largest frontier models, unlocking real-time intelligence for edge devices." Among disclosed adopters, Alphabet's drone delivery unit Wing is evaluating the module, while home-cleaning robot company Matic has already adopted it to run conversation, gesture detection, and semantic fine-grained mapping locally; Cognex and Doosan Bobcat are also on the list. Aptiv separately announced the same day that its production-grade platform supports the module. Source: Unite.AI source

XPeng G9L debuts second-generation VLA, with X-Foresight predicting 6-second trajectories · autonomy

An XPeng vice president announced today that the G9L will be the first vehicle to feature a new version equipped with second-generation VLA, with its accompanying X-Foresight world-prediction model able to project environmental motion trajectories 6 seconds into the future. He Xiaopeng said earlier on the earnings call that the second-generation VLA model will get its first major upgrade at the end of August, with the on-device model's parameter count increasing 3.5x. The vehicle opened for pre-order on August 11 at a pre-sale price of RMB 259,800, available in super range-extender and pure-electric variants, with 800V high-voltage ultra-fast charging across the lineup. A Physical AI sharing and experience event is scheduled for August 27 at 15:00. Source: Sina Technology source

5-year-old girl loses 4 teeth after being kicked by a performance robot; operator confirms it was a Unitree product · humanoid

At a science museum in Xiangyang, Hubei, a 5-year-old girl watching a robot performance was struck in the face by the robot's raised leg, losing 4 teeth and requiring 3 stitches to her face. The incident occurred in November 2025; the museum said the operator "failed to anticipate the safe distance and misoperated the robot." China Newsweek confirmed with the operating company, Andar Innovation Exhibition, whose marketing manager stated explicitly that the robot was a Unitree product. Similar incidents have occurred elsewhere: in June 2026, a humanoid robot performing martial arts in Urumqi struck a boy in the abdomen with a spinning kick, and in February an approximately 1.6-meter-tall robot at Chengdu's CapitaMall deviated from its preset path and knocked over an elderly person; neither site had physical barriers separating the robots from the public. The market itself is expanding — iiMedia Research data shows China's robot rental market was worth about RMB 1 billion in 2025 and is projected to exceed RMB 10 billion in 2026, with Orient Securities giving a similar-magnitude forecast for the performance-robot rental market specifically. One amusement-park operator told the magazine, "the barrier to entry is genuinely low — if you can afford one, you can run this business"; the venue keeps 4 dedicated operators on staff and requires crowds to be cleared before major movements, but acknowledged the industry has no unified standards — "all of this relies on self-discipline." The family has repeatedly requested the operator's qualifications and training records, none of which have been provided to date. Source: iFeng source

Qiongchu Intelligent releases Noe-0, an entirely teleoperation-data-free pipeline · world-model ⚠️ company account

Using no teleoperation data at all in either pretraining or post-training is this world-action model's most aggressive design choice. The team used its self-developed RoboPocket collection device, letting operators perform tasks directly by hand in real homes and stores, gathering self-collected data spanning over 50 cities, several hundred thousand hours, and hundreds of thousands of task types. One of the reasons the team gives is practical: completing the same task via teleoperation typically takes 3 to 5 times longer than direct manual operation, and tends to produce slow, mechanical motions. Architecturally, the policy backbone is swapped from a language model to a video world model, using video prediction as the core high-level learning objective; the team says that even when there is a mismatch between the collection setup and the deployment robot body, pixel prediction still meaningfully improves cross-embodiment transfer. Technical details were published on the team's research preview blog; weights have not been open-sourced. Source: Sina Technology source

Kodiak puts AMD EPYC chips in its seventh-generation driverless truck · autonomy

In the seventh-generation Kodiak Driver, an EPYC CPU handles fusion and preprocessing of lidar, camera, and radar data while also supporting path planning. AMD's senior director of automotive markets, Wayne Lyons, said the key CPU workloads in autonomous driving are latency-sensitive and hard to parallelize. Kodiak founder and CEO Don Burnette said using off-the-shelf commercial hardware lets the company focus resources on building trucks that can be mass-produced. The company's driverless fleet in the Permian Basin now operates with no one in the cab. Source: The Manila Times source

Hardware · Supply Chain

· Standardized robot battery: JD.com launched a standardized battery series at WRC 2026, with cells supplied by BAK Battery, using unified interfaces and dimensions for plug-and-play compatibility across humanoid, industrial, and wheeled robot forms; BAK is also pushing a "one cell, three variants" strategy for humanoids, pairing a mature 21700 cell with a unified physical/electrical architecture, then offering burst-output (running, jumping, load-bearing climbing), endurance (commercial service), and balanced (home and industrial collaboration) variants source

· Silicon carbide for humanoids: UBTECH (Chinese humanoid robot maker) signed a strategic partnership with BASiC Semiconductor around SiC power devices, for joint research and validation on power-management and motion-control efficiency and endurance improvements in robots; BASiC Semiconductor plans to deploy UBTECH humanoid robots on its own silicon-carbide production lines for wafer fabrication and module packaging source

FutureX · Physical AI Daily — Issue 100 (08/26)

Shawn — Tue, 25 Aug 2026 15:00:09 +0000

Today's Highlights

· Draft amendment to China's Road Traffic Safety Law adds a dedicated chapter on autonomous driving; violations to be handled by manufacturers and importers

· Skild AI lands $1.4 billion Series C at a valuation over $14 billion, led by SoftBank

· Gatik closes $200 million Series D, has completed 85,000 fully driverless deliveries

· Tiangong robot runs 100m in 8.86 seconds, half a second faster than its opening-ceremony record

· New FCC rule: robots assembled in the US must exceed 65% domestic component content by 2028

Research Progress

DELE-w0.5: World models don't need to render video first · world-model

Video generation is an unnecessary intermediate objective for world action models — this paper turns that claim into an experiment. DELE-w0.5 skips frame-by-frame prediction and infers action sequences directly from compressed future latent states, on the reasoning that what a robot needs is what the world looks like after the action completes, not what it looks like at every intermediate moment. Across four long-horizon manipulation tasks and 480 real-robot trials, overall task success reached 62.5, 47.5 percentage points above the strongest baseline, while cutting both training and inference cost.

Fenghao Lei et al. · arXiv 2608.22067 source

GeoWAM: A driving world model that predicts point clouds, not images · autonomy

Pixels entangle geometry, motion and lighting together, yet driving world models must infer 3D transformations from 2D images. GeoWAM replaces the state space so the pretraining target becomes the future scene's point-cloud geometry rather than future images, with an action head predicting the ego-vehicle trajectory conditioned on this geometric dynamics. Both open-loop and closed-loop evaluations show driving policies built on geometric world modeling clearly outperform image-based counterparts.

Yiren Lu et al. · arXiv 2608.23486 source

LD4WAM: Feeding 5,000 hours of human video into a world action model · world-model

Human video is cheap and diverse, but it can only teach pixel-level dynamics that aren't directly executable; action retargeting, in turn, leaves a large cross-embodiment visual gap. LD4WAM proposes motion-aligned latent dynamics as an intermediate representation — preserving full future video generation on one side, while using learnable queries to distill this dynamics layer to condition the action. The pretraining corpus is a self-built dataset of over 5,000 hours of human and robot data, tested successfully on the RoboTwin simulator as well as on real grippers and dexterous hands, holding up even on unseen objects and backgrounds.

Zhenhao Shen et al. · arXiv 2608.22403 source

Consistency loss placed in the wrong spot can hurt — and exposes inflated published numbers · world-model

Change the camera angle and world action models fall apart. This paper offers a near-counterintuitive finding: applying cross-view consistency loss to outputs that legitimately vary with viewpoint — like predicted future frames — is provably harmful; the correct approach is to constrain only the viewpoint-invariant parts, such as action chunks, future proprioception, and value. The method requires no camera calibration, extrinsics, or depth; on held-out viewpoints outside the training view envelope, closed-loop success is 12.2 percentage points above the paired control, with no gain within the envelope. A companion cross-backbone audit finds that previously published camera-robustness numbers were confounded by wrist-camera pose stability.

Bingqi Huang et al. · arXiv 2608.21402 source

The Imitator Game: Nine SOTA models collectively collapse at level four · benchmark

Humans imitate intent; robots learn trajectories. This benchmark splits that gap into four levels, L0 through L3, progressively widening the distance between human demonstrations and the robot's own scenario. Nine SOTA models perform steadily from L0 to L2, then collapse across the board at L3, where the bottleneck is functional substitution — using a different object's affordance to achieve the same intent. The accompanying IG-10K, currently the largest environment-aligned human-robot paired dataset, contains over 20,000 paired clips across 50-plus tasks and 6 domains, spanning both real robots and simulation; all models score a zero-shot success rate below 13% on unseen tasks, while fine-tuning on just 10 paired demonstrations yields significant gains that scale with pretraining size.

Xunzhe Zhou et al. · arXiv 2608.22301 source

Adding one retrieved sentence to the prompt drops VLA success rate from 92.47% to 3.00% · vla

The authors call this phenomenon prompt-form collapse: changing the form of an instruction alone is enough to dominate execution, and whether the appended semantics are meaningful is secondary — both meaningful text and length-matched nonsense text fail identically across all 500 states. TOWN-VLA's countermeasure adds an authorization gate to the prompt: only compact instructions meeting fixed compatibility rules are allowed to rewrite the policy input, otherwise the original prompt is restored byte-for-byte. On LIBERO-Plus, with 10,030 episodes per method, success rose from 69.5% to 73.1%; on a real PiPER robotic arm paired with a frozen π0.5 checkpoint, success rose from 52.7% to 78.7% across 150 trials.

Zhiruo Zhou et al. · arXiv 2608.23224 source

GOLEM: A humanoid disassembles retired EV battery packs, open source, with a capability ladder · manipulation

Disassembling scrapped battery packs is tedious and dangerous work, currently done almost entirely by hand. GOLEM decomposes walking, manipulation, dynamic balance, navigation, and spatial memory on a Unitree H1-2 humanoid into independent, interface-abstracted modules, wrapped in a Dockerized ROS 2 setup, with MuJoCo and IsaacLab digital twins matching the real robot's interface, all open source. The real value is its step-by-step evaluation ladder: grasping already-loosened fasteners from a real Hyundai Ioniq 5 battery pack, success reaches 97% while suspended, drops to 87% when freestanding, and falls to just 37% once pose deviation from navigation is layered on.

Max Conway et al. · arXiv 2608.21550 source

A single photo of a clinic room generates a simulated scene robots can be tested in · benchmark

Embodied AI needs to be tested in the clinical environments it will actually work in, but building a realistic, robot-ready site is expensive and hard to scale. Researchers validated the approach on 39 ophthalmology clinic scenes, converting a single photo into an editable environment ready for direct use in a simulator, then adding equipment meshes, collision proxies, and semantic anchors to turn visual reconstruction into a scene where contact can be computed. Three robot morphologies show clearly different reachability and contact feasibility on the same set of task goals; small translations and rotations in equipment position change the contact margin in ways that visual similarity alone cannot reveal.

Xinyuan Wu et al. · arXiv 2608.21416 source

InstructMove: A manipulation benchmark that fails you if you don't read the instruction · benchmark

Most manipulation benchmarks can't tell whether a policy is actually listening to instructions, since the target object is often the most visually salient one, or physically only one action is feasible — so guessing still works. InstructMove requires multiple actions in the scene to be both visually and physically feasible, with only one consistent with the language, then splits instruction-following into four categories: category recognition, attribute discrimination, spatial reasoning, and compositional pick-and-place. Experiments show it can diagnose visual shortcuts a policy is taking, and its simulated data also improves real-robot instruction-following performance; the code is hosted in the Horizon Robotics RoboOrchardSim repository.

Mengao Zhao et al. · arXiv 2608.22990 source

Other papers today: WorldToken fuses multi-view images, proprioception and task conditioning into a single world token at each timestep; an 85.3M-parameter policy reaches 59.45% closed-loop success across 23 RoboCasa tasks (arXiv 2608.22591 source); WAM-OPD uses on-policy distillation to repair world action model students whose capability degraded after acceleration (arXiv 2608.22364 source); DreamMimic uses a world model to teach humanoid whole-body visual loco-manipulation (arXiv 2608.22278 source); DELTA uses a deformable elevation attention encoder to decode sparse terrain foot placement for quadrupeds (arXiv 2608.22033 source); CounterAlign adds counterfactual negative supervision to VLA behavior cloning that otherwise only has positive samples (arXiv 2608.21740 source); GuardianBench fixes the scene and varies only the instruction to test embodied models' latent situational risk (arXiv 2608.21928 source); MomADv2 suppresses inconsistent temporal memory when driving instructions change (arXiv 2608.23405 source); OpenSCvx is an open-source Python framework for nonlinear trajectory optimization (arXiv 2608.21631 source); Astrobee grayscale images plus 3D Gaussian Splatting reconstruct the ISS's Japanese experiment module (arXiv 2608.21685 source).

Open Source · Tools · Benchmarks

· RoboColiseum: a standardized embodied-AI simulation evaluation platform launched by a Shanghai team, with task suites and leaderboards across four capability dimensions; the team says the sim-to-real gap for the same model is under 10%, and hundreds of teams have already run models on it during closed testing source

· HOMIE Gen2: a head-mounted capture rig from Singapore's Ropedia, with four cameras, spatial audio and inertial data hardware-synchronized to within 50 microseconds; recordings include camera pose, depth, hand and body keypoints, and task annotations — the company says it is the first first-person robot training capture device with a 360-degree field of view source

Funding & Deals

Skild AI ｜ Series C ｜ $1.4 billion ｜ post-money valuation over $14 billion · embodied

Led by SoftBank, with participation from Nvidia, Bezos Expeditions (Jeff Bezos's investment vehicle), Macquarie Capital, and 1789 Capital, plus Samsung, LG and Schneider Electric joining as strategic shareholders. The company, three years old, has raised over $2 billion total. Skild doesn't build hardware — it sells Skild Brain, software that drives quadrupeds, humanoids, desktop robotic arms, and mobile manipulation robots; ABB, Universal Robots, and Mobile Industrial Robots have already embedded it in their machines. The software currently runs on hundreds of robots across factories, data centers, and logistics hubs, including deployments at an Nvidia factory in Houston and a pilot at LaGuardia Airport. In under half of 2025, Skild recorded roughly $30 million in revenue. Co-founder and president Abhinav Gupta, who left Meta's FAIR robotics lab to found the company with Carnegie Mellon professor Deepak Pathak, said "this is the beginning of the GPT moment, and it's going to happen over the next year to a year and a half" — the bet is on factories and warehouses, not homes.Source: Dealroom source

Gatik ｜ Series D ｜ $200 million · autonomy

Co-led by the Qatar Investment Authority and Koch Disruptive Technologies, with participation from Millennium Management, ARK Invest, and Intact Private Capital. Gatik runs short- and mid-haul driverless freight between distribution centers and stores, with routes across Texas, Arkansas, Arizona and Canada, serving customers including Walmart, Tyson Foods, and Canadian grocery chain Loblaws. The company says it has completed 85,000 fully driverless deliveries, with over $600 million in contracted revenue and a 99% on-time rate. In January, CEO Gautam Narang told Forbes the fleet would grow from just over a dozen vans to hundreds of vans by the end of 2026.Source: Forbes source

Generalist AI ｜ new round ｜ $200 million · embodied

Led by 8VC, with participation from some existing investors; valuation undisclosed, coming just two months after its $400 million round in June. About a week earlier, the company released Gen-1.5, aimed at compressing the setup time for factory automation workflows: a user wearing sensors or using the robot's own camera demonstrates a task once, and the model learns the new task — across 10 sample tasks, average completion rate was 59% from a single demonstration (previously reported).Source: Axios (via Digital Today) source

Airbound ｜ Series A ｜ $37 million · adjacent

Led by Greenoaks, with participation from DoorDash, Lightspeed, Lachy Groom, and Humba Ventures, less than a year after an $8.65 million seed round, bringing total funding to nearly $50 million. This Indian company builds tail-sitter VTOL drones designed so the aircraft weighs less than its payload: the current model, TRT, weighs about 3.3 pounds and carries 2.2 pounds, while a model in development weighs about 6.6 pounds and carries up to 11 pounds. Over 13,000 autonomous flights have been completed in Bangalore and Guntur, including more than a thousand delivering lab samples for the Narayana Health hospital network — a 2.5-mile trip taking about 7 minutes, versus 3 to 5 hours by two-wheeler on the ground including batching and wait time.Source: TechCrunch source

General Intuition ｜ new round ｜ valuation around $6 billion · world-model

Following reports on the 23rd that talks were underway, the investors have been revealed as Valor and Point72, as the company simultaneously shifts its focus toward robotics.Source: TechCrunch source

NEC × Dexmate ｜ strategic investment ｜ amount undisclosed · embodied

NEC's Orchestrating Future Fund has invested in US robotics company Dexmate, maker of the VEGA humanoid robot.Source: NEC source

Commercialization & Deployment

Galbot S1 (Chinese robotics startup) runs 7x24 on a CATL production line · embodied

Galbot S1, with a maximum dual-arm payload of 50 kg, is now running 7x24 on a CATL production line, handling depalletizing, transport, and palletizing of materials. At the same WRC event, Galbot also unveiled its humanoid robot Xingzai and the embodied foundation model that drives it, AstraBrain, with the company saying the same general-purpose brain can control bipedal, wheeled, and heavy-load robot bodies simultaneously. A booth demo and stable running on a production line are two different things — the former's reliability cannot be directly extrapolated to the latter.⚠️ Vendor claimSource: Gasgoo source

JAKA Robotics' (Chinese robotics manufacturer) embodied-AI revenue hits 45 million yuan in H1, up more than 20-fold year over year · industrial

JAKA Robotics reported H1 revenue of about 320 million yuan, up 106.6% year over year, with embodied-AI business revenue topping 45 million yuan, up more than 20-fold year over year and accounting for nearly 15% of total revenue; the company has 231 embodied-AI customers to date. R&D spending was about 100 million yuan, up 148.4% year over year, widening overall losses accordingly. The company says its self-developed DobotWAM model achieved a 99.25% average success rate on LIBERO, higher than mainstream published models as of results available through 2026.⚠️ Vendor claimSource: Sohu source

Keenon Robotics' KOM 3.0 packs a latent-space world model into service robots · embodied

The prior generation, KOM 2.0, used a fast-slow dual system: the slow system, K-Mind, handled task understanding and planning, while the fast system, K-Act, handled fine-grained motion generation. KOM 3.0 builds on this by integrating a latent-space world model, which Keenon calls the world's first VLA architecture for the service industry to incorporate a latent-space world model. At the WRC venue, the company set up four scenarios — a coffee kiosk, a dessert shop, a retail store, and a laundromat — where its XMAN-R1 humanoid, acting as a guest barista, autonomously completed the full workflow from cup retrieval to extraction, while two humanoids in the laundromat coordinated washing and folding under unified scheduling. COO Wan Bin said cumulative shipments have exceeded 100,000 units, with business covering more than 60 countries and regions.⚠️ Vendor claimSource: Gasgoo source

Amazon planning fully automated delivery stations "Tetromino"; company says document figures are inaccurate · industrial

An internal planning document from last month, seen by Business Insider, states that Amazon plans to invest $103 million in its first Tetromino pilot in 2028, build five sites in 2029 at about $85 million each, add ten more in 2030, and reach cumulative investment of over $530 million by 2029. The document says the design's package-processing rate is about 2.5 times that of existing delivery stations. An Amazon spokesperson told BI this is an "early-stage concept" and that the specific financial figures and roadmap in the document are "inaccurate and do not reflect our current plans." The delivery station is the last step before packages reach drivers, and put-away and loading are widely regarded in the industry as the hardest and most expensive parts to automate. The document mentions that one key technology may come from startup Boxbot, whose pallet storage-and-retrieval system claims to speed up loading by up to tenfold. On its July earnings call, Amazon said the number of robotic arms would more than double in 2026.Source: Business Insider source

Waymo selects Munich as its first EU robotaxi city · autonomy

The Financial Times reports Waymo has designated Munich as its first EU launch city, having already registered a German subsidiary, Waymo Germany GmbH, in June. Separate reports say it has also set up a presence in Madrid to prepare for autonomous-taxi testing. On the same day, Amazon's Zoox expanded its steering-wheel-free robotaxi from Las Vegas to additional cities including San Francisco (previously reported as having started charging fares).Source: Financial Times source; Olive Press source; NBC News source

UBTECH and Singtel sign memorandum of understanding · humanoid

The two companies signed an MOU at the 2026 World Robot Conference aimed at combining embodied-AI robots with telecom service capabilities, to advance robot deployment in Asia-Pacific enterprise operations. The memorandum did not include a dollar amount, delivery volume, or timeline.⚠️ Planning stageSource: Beijing Business Today source

Industry Developments

Draft amendment to China's Road Traffic Safety Law adds dedicated autonomous-driving chapter; violations handled by manufacturers and importers · autonomy

The draft amendment to the Road Traffic Safety Law was submitted on the 25th for its first reading at the 24th session of the Standing Committee of the 14th National People's Congress, running 9 chapters and 170 articles in total. The autonomous-driving section contains four articles: self-driving vehicles must pass a road-traffic-rule compliance test and be registered in accordance with the law before they can operate on public roads; traffic violations occurring while the autonomous driving system is operating are to be handled by the manufacturer or importer, with specific procedures to be drafted by China's public security authorities under the State Council; a compulsory third-party motor vehicle traffic accident liability insurance system will apply to self-driving vehicles, with rules to be set by the State Council; and manufacturers and importers must ensure road safety, cybersecurity and data security, and may not falsely or exaggeratedly advertise autonomous-driving functions. The current Road Traffic Safety Law took effect in 2004, and its three amendments in 2007, 2011 and 2021 each touched only isolated provisions. Shifting the party responsible for violations from the driver to the manufacturer or importer would directly rewrite the compliance cost structure for autonomous-driving companies operating in China. The draft also includes provisions capping e-bike speed at 20 km/h, banning "distracted riding," and regulating so-called "roving mob" group cycling.Source: China News Service source

FCC adds advanced mobile robots to Covered List; US-made robots must exceed 65% domestic content by 2028 · hardware

The US Federal Communications Commission added advanced mobile robots to its Covered List in July. Per the FCC website's Q&A page, new advanced robots assembled in the US must have domestic component value exceeding 65% by 2028, rising above 75% by 2029. Regulators have yet to clarify how exemptions will work. Jeff Burnstein, president of A3, an industry association representing nearly 1,500 robotics and automation companies, said "everyone is trying to figure out what this means for them and what the impact will be," calling the regulatory environment "a day-to-day thing right now." Peter Cowhey, professor emeritus at UC San Diego and a former FCC official, told Nikkei that the rule's intent is to diversify the supply chain and bring some capacity back onshore, but "the problem is there's no concrete mechanism to get from here to there, to swap Chinese products for American ones" — calling it "an incoherent policy of forced import substitution."Source: Nikkei Asia source

Tiangong robot clocks 8.86 seconds in the 100m · humanoid

In the first heat of the men's large-class 100m at the second World Humanoid Robot Games, Tiangong, built by the Beijing Innovation Center for Humanoid Robotics, ran 8.86 seconds, shaving over half a second off the 9.39-second record it had just set at the opening ceremony. Reuters covered the same event under the headline that a Chinese robot ran under 9 seconds in Beijing. Elsewhere at the games, things were less dignified: a humanoid collided with the judges' table during a weightlifting event and was carried off on a stretcher.Source: Sichuan Observer source; Reuters source; EssentiallySports source

XPeng's second-generation VLA upgraded to 6.3.0, on-device parameter count up 3.5x · autonomy

The new version increases the on-device model's parameter count to 3.5 times the previous version and improves end-to-end response speed by 300%; XPeng says this scale is 15 times that of mainstream VLA models in the industry. The launch vehicle, the G9L, uses XPeng's Turing AI chip — one chip for the Max trim, two for Ultra SE and Ultra. Version 6.3.0 brings some robotaxi-grade L4 experiences into production vehicles for the first time, including smart standing starts, precise curbside stops, and autonomous parking-spot search in campus settings; the system supports over 30 seconds of temporal reasoning and can anticipate scenarios up to 6 seconds ahead. The VLM has also been added to the cabin, supporting vague-destination navigation and voice-guided nearby parking. The G9L has accumulated over 6.74 million km of R&D testing mileage across 26 countries and regions, with rollout starting in late August.⚠️ Vendor claimSource: Netcarshow source

He Xiaopeng: IRON to enter mass production by year-end, formal launch in 2027 · humanoid

At the Q2 earnings call, XPeng CEO He Xiaopeng gave a timeline: mass production begins by the end of this year, starting with rollout at XPeng's own stores and campuses; the IRON humanoid formally launches in 2027, aiming for large-scale delivery to external retail and service-industry customers in China and overseas, with monthly production capacity scalable to several thousand units next year depending on market demand. He said IRON has 76 degrees of freedom across its full body, with 21 in each dexterous hand, and that the robotics R&D effort has spanned eight years, with the technical challenge of a high-end general-purpose humanoid at least 20 times greater than that of a smart car. He also predicted that once IRON launches, per-unit lifetime revenue and gross-profit contribution will be far higher than the current per-vehicle average price and margin in XPeng's auto business. The humanoid unit completed its first funding round last week, raising over $900 million at a post-money valuation above $6.3 billion (previously reported).⚠️ Executive remarksSource: Kuaikeji (via Sina Tech) source

NIO's autonomous-driving chief Ren Shaoqing founds a physical-AI foundation model company · world-model

At an autonomous-driving all-staff meeting on August 24, NIO CEO William Li announced that Ren Shaoqing has founded an independent company focused on physical-AI foundation models and embodied intelligence; NIO will make a strategic investment and collaborate with the new company, while Ren continues to head NIO's autonomous-driving business. The new company has been registered, though its name and funding details have not been disclosed. Ren is one of the four authors of ResNet and one of the creators of Faster R-CNN; after joining NIO he led the mass-production delivery of NWM, described as China's first large-scale, forward-simulating world model for autonomous driving. 36Kr separately reported the new company achieved unicorn status upon registration, which does not square with reports that its funding remains undisclosed.Source: LatePost (via ThePaper/Zhidongxi) source; 36Kr source

South Korea launches K-Moonshot AI humanoid initiative, aims to name a national-team platform within three years · humanoid

South Korea's Ministry of Science and ICT held a meeting on the 25th to gather input from industry, academia and research institutes on a draft "K-Moonshot AI Humanoid Roadmap," with the goal of producing a world-class national-team K-humanoid platform within three years, localizing AI models and key components, and running the effort from core technology development through demonstration and mass production. Also announced was the Next-X Humanoid Challenge, which will evaluate humanoid robots developed in South Korea and select the best performers for support, with the challenge ultimately choosing a national-team-level platform starting in 2027. The initiative's overall lead, Yeo Jun-gu, said that after AI, the next wave is physical intelligence — systems that can judge and move autonomously in the physical world — with humanoids as its core platform. The roadmap will incorporate feedback from this session before going to a meeting of science-related cabinet ministers, with a final version due next month.Source: Digital Today source

Shenzhen industrial robot output up 45.5% year over year, January–July · industrial

According to Shenzhen Municipal Bureau of Statistics data released on the 24th, industrial value-added output above designated size in the city rose 9.2% year over year for January–July, 0.5 percentage points faster than the first half of the year, with specialized equipment manufacturing up 14.4% and general equipment manufacturing up 14.3%. Among high-tech products, output of 3D printing equipment, industrial robots, and lithium-ion battery products rose 62.3%, 45.5%, and 27.1% respectively.Source: Cailianshe (via Sina Finance) source

Hardware · Supply Chain

· Actuators: A report from Gavekal Technologies dated August 10 says actuators typically account for 40%–60% of a humanoid robot's hardware BOM, with batteries, lidar and motor controllers also mostly sourced from China — cited as the direct reason the US domestic-content threshold will be hard to meet (via Nikkei Asia) source

· Dexterous hands: Forbes visited Beijing-based Linkerbot (Chinese dexterous-hand startup), which the company says was valued at $3 billion in an April Series B+ round backed by Ant Group, HSG, and CICC Capital; a report from GGII estimates Chinese dexterous-hand shipments will reach about 70,000 units this year, up nearly 300% from 2025, expanding to 430,000 units annually by 2030; Tianfeng Securities estimates over twenty Chinese companies were already making robotic hands as of 2025 source

· AAEON robot development platform: a robot development platform based on Intel Core processors has launched, targeting edge-side compute configurations source