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
Top comments (0)