DEV Community: Shawn

FutureX · Physical AI Daily — Issue 69 (07/26)

Shawn — Sat, 25 Jul 2026 14:43:13 +0000

Today's Highlights

· Waymo notified Uber it will launch its own app in Austin and Atlanta in January 2028, ending exclusivity in both markets

· Hyundai Motor announced its transformation into a "Physical AI Solutions Company," co-building a robotics reference platform with Nvidia

· Dongfeng Aeolus (Chinese automaker) × Hello Inc. (Chinese mobility platform) L4 robotaxis roll off the production line, with plans to deploy 10,000 units and expand to no fewer than 10 cities within two years

· Dexterous hands lifting 5 kg break after only about 50 cycles; even top products last just 300,000-500,000 cycles

· AGIBOT's (Chinese humanoid robotics startup) Hong Kong IPO could land as early as August, with cornerstone investors targeting a valuation of HK$40-50 billion

I. Research Progress

Nankai University's "one chip, three functions" ferroelectric transistor: squeezing sensing, computation, and activation into a single device · perception

Embodied vision systems have long been constrained by sensors, memory, computation, and activation circuits each operating in isolation — data must repeatedly shuttle between modules and undergo multiple analog-to-digital conversions, keeping both energy consumption and latency stubbornly high. A team led by Chen Xudong, Tian Jianguo, and Chen Shuqi at Nankai University, in collaboration with Kang Lixing's team at the Suzhou Institute of Nano-Tech and Nano-Bionics (Chinese Academy of Sciences), used a dual split-control-gate structure to modulate local polarization in CuInP2S₆, programming p-n/n-p/p-p/n-n homojunctions into an ambipolar WSe₂ channel. This lets a single ferroelectric transistor switch on demand between three modes — in-sensor computing, in-memory computing, and nonlinear activation — with 33 optical weight states, 32 multi-level conductance states, and ReLU/Sigmoid-type response times of about 8 ns. The team built end-to-end neuromorphic vision hardware on this basis, converting classification results directly into left-turn/right-turn/stop commands for a mobile robot, achieving 95% MNIST recognition accuracy, roughly 98% system-level simulation accuracy on traffic video processing, and 0.99 structural similarity between the full-hardware autoencoder reconstruction and the software benchmark. The work was published in Science Advances.

Li Yuan, Zhang Zhicheng et al. (Nankai University · Suzhou Institute of Nano-Tech and Nano-Bionics, CAS) · Science Advances 10.1126/sciadv.aed9436 · via: Nankai University Institute for Interdisciplinary Sciences source (WeChat, CN)

Jetson-PI: turning VLA from "demo-capable" into "continuously operable" on onboard compute · vla

Running π0.5 directly on Jetson Orin yields an inference time of about 1.4 seconds and a control frequency of only around 0.7 Hz, leaving the robot sluggish with visible pauses between action chunks — this is the most concrete wall facing edge-device VLA deployment. Peking University, AIRS, and PrimeBot Research Institute propose "look-ahead aligned asynchronous correction": rather than fusing old and new trajectories, they train a lightweight module of about 40M parameters (roughly 1% of the full VLA) to directly predict the future VLM environment representation after the current action completes, letting the action expert generate actions starting from that future moment — sidestepping the perception-execution misalignment inherent to asynchronous inference. The system also schedules by confidence: it skips the VLM and runs only the action expert during steady states, and only "looks again" when confidence drops at critical steps like grasping or placement. Without sacrificing accuracy, this raises the control frequency on Jetson Orin to 6.06 Hz, 8.66 times faster than native PyTorch inference.

Peking University · AIRS · PrimeBot Research Institute · arXiv 2607.12659 source · via: Sina Mobile source

Open Source · Tools · Benchmarks

· Jetson-PI: the Peking University team has open-sourced both the asynchronous inference code and the edge inference engine (PKU-SEC-Lab), targeting low-power onboard devices like the Jetson Orin/Thor source

· DexLink / SyLink Hand: material costs disclosed in the papers have dropped to the $400 range — DexLink drives 20 joints with 16 actuators in a 320-gram hand, completing all 33 Feix grasp types; SyLink goes further, driving 19 joints with just 11 actuators. The reference point is the LEAP Hand's sub-$2,000 cost in 2023, while the LEAP Hand V2's metal-palm durability variant actually rose to about $4,000 — the cost conversation is shifting from "how much does one hand cost" to tolerances, calibration, and lifespan source (WeChat, CN)

· ORCA Hand: material cost under 2,000 Swiss francs, with a publicly verified continuous run of 7 hours 17 minutes completing about 2,000 grasp-and-release cycles without tendon loosening or breakage — pushing benchmarking from "can it do it once" to "can it repeat it unattended" source (WeChat, CN)

II. Funding and Deals

AGIBOT (Chinese humanoid robotics startup) | Hong Kong IPO | Cornerstone target valuation HK$40-50 billion · humanoid ⚠️ Reported terms

Following the formal launch of its Hong Kong listing process on July 24, AGIBOT could complete its Hong Kong IPO as early as this August; cornerstone investors have set a target valuation of HK$40-50 billion (roughly RMB 34.6-43.3 billion), while AGIBOT itself had hoped to push the target valuation to around HK$80 billion. For reference, the company's most recently disclosed post-money valuation as of July 2026 was RMB 15 billion, with 2025 revenue exceeding RMB 1.05 billion — nearly a 20-fold increase year over year. According to Omdia, AGIBOT shipped over 5,100 units of general-purpose humanoid robots in 2025, roughly 39% of the global market, while rival Unitree Robotics (Chinese quadruped/humanoid robot maker) officially disclosed shipments exceeding 5,500 units over the same period — the discrepancy between the two figures illustrates that "world's number one" depends on how the statistics are scoped.Source: Zui Investment Bank source (WeChat, CN)

Jingling Zhikang (Chinese wearable prosthetics startup) | Angel round | Nearly RMB 100 million · embodied

Led by Longpan Investment, with participation from Jiangxi Financial Holdings, Xinhuadu Chairman Ni Guotao, and USTC-Silicon Valley (Hefei). This funding isn't targeting a dexterous hand mounted on a robot's end-effector — it's targeting one worn on a human hand. Hefei-based Jingling Zhikang makes a wearable bionic hand weighing about 370 grams that reads the wearer's movement intent to drive grasping motions. Price is the real battle it's fighting: imported high-end smart bionic hands typically cost RMB 300,000-500,000, and comparable Chinese products generally exceed RMB 100,000. The company — which completed a strategic acquisition of Lingxin Qiaoshou this past May — proposes first cutting the price to RMB 30,000-50,000, aiming to reach the RMB 10,000 price point within three years, with planned production of 500 units in 2026, 2,000 in 2027, and a push toward 10,000 in 2028. The company states its latest tests show a lifespan exceeding one million cycles⚠️ vendor claim.Source: Touyingli source (WeChat, CN)

Acrab (Singapore) | Cumulative funding | $350 million · hardware

Led by a $50 million investment from Huangpu River Capital (Chinese venture capital firm), with Temasek's Vertex making a heavy bet through Vertex Ventures SEA & India and Vertex Growth. Founder Ken Phua previously served as Arm's VP of Global Technology Strategy and Acquisitions and co-CEO of Arm China; the company is betting that "the CPU is becoming the center of the AI era again." Its first chip, the 5nm GΞLIX 1, has been announced — a heterogeneous architecture integrating a 20-core Arm CPU, a 3 TFLOPS GPU, and a dedicated NPU, delivering 650 TOPS of total AI compute. The company claims that in a Gemma 26B A4B configuration, prefill throughput reaches 1,416.8 tokens/s — 7.5 times that of an M4 Pro Mac Mini⚠️ vendor claim. Lead investor Huangpu River Capital has previously backed Montage Technology, Horizon Robotics, Galbot (Chinese humanoid robotics startup), and Paxini (Chinese tactile sensing startup).Source: LeiDi · Sina Finance source

Yisuo Intelligence | Seed round | Tens of millions of RMB · world-model

Exclusively invested by Zhengxuan Investment; the company simultaneously disclosed a term sheet for nearly RMB 100 million. As world models collectively pivot toward robotics and autonomous driving, Yisuo Intelligence is targeting scientific world models instead: its Iso-SWM engine + Iso-Sciland research workspace + Iso-LabOS experimental feedback system run a "predict-experiment-calibrate-repredict" wet-dry closed loop, starting in life sciences before expanding into materials, energy, chemical processes, and semiconductors. The reference point is Isomorphic Labs (Alphabet subsidiary)'s $2.1 billion funding round completed this May — this segment has clearer customers and willingness to pay, but requires a combination of AI algorithms, domain knowledge, and wet-lab engineering capability.Source: Rongleme source (WeChat, CN)

III. Commercialization and Deployment

Dongfeng Aeolus × Hello Inc.: L4 robotaxi rolls off production line, plans to deploy 10,000 units · autonomy

Following the delivery of Hello Inc.'s first mass-produced robotaxi, the HR1, yesterday, the vehicle officially rolled off the line at the Dongfeng Nissan Huadu plant on July 24 — the project was jointly built by Dongfeng Aeolus and Guangzhou Zaofu Intelligence (Hello Inc.'s inclusive mobility arm), based on the Aeolus VX6 electric SUV, with plans to deploy 10,000 units at scale. The first batch will run routine demonstration operations in Guangzhou and Shenzhen, with plans to expand to no fewer than 10 cities nationwide between 2026 and 2027. What's genuinely worth noting is the model itself: "automaker mass production + mobility platform's in-house autonomous driving" — shifting robotaxis from the retrofit-vehicle route to a product coming straight off an automaker's production line. The VX6's cabin volume can accommodate lidar and onboard computing units, and its economics have already been validated through high-intensity ride-hailing operations.Source: Caijing 21 source

FAW Jiefang delivers 73 L2 navigation-assisted heavy trucks to Annto Logistics, targeting "two drivers to one" · autonomy

On July 24 in Changshan, Zhejiang, FAW Jiefang delivered the first batch of 133 new trucks to Annto Logistics, of which 73 units are J6E L2 navigation-assisted heavy trucks entering operation on the national trunk-line transport network. The company states the model, equipped with its self-developed "vehicle + assisted driving" system, can achieve fuel savings of 3% to 5% and enable a shift from "two drivers to one" on mid-to-short-haul trunk routes of 500-1,000 kilometers — this is a direct cost line item for trunk-line logistics, not autonomous driving. As of June 2026, Annto operated over 11,000 company-owned trucks and trailers, with network coverage reaching about 99.7% of counties nationwide, and 1,376 FAW Jiefang trucks now in operation on its network.Source: Shanghai Securities News · Sina Finance source

South Korean Navy uses a humanoid robot as a ship's helmsman for the first time, Pibot passes phase-one testing · humanoid ⚠️ Simulator stage

The South Korean Navy, in partnership with KAIST (Professor Shim Hyun-chul's team), placed the humanoid robot Pibot in a helmsman role for the first time in a combat exercise: when the officer of the watch ordered "left rudder 5 degrees, course 330," the robot repeated the command, turned the wheel, and — once the simulated ship stabilized — reported "steady on course 330," completing the full repeat-back, steer, and report cycle a human helmsman would follow, with command comprehension driven by a large language model. The test spans four phases, with this being the first-phase land-based ship-handling simulator; subsequent phases will cover in-port mooring, daytime sea navigation, and continuous day-night navigation, with researchers analyzing time-to-completion and steering precision from command to execution. The South Korean Navy's motivation is stated plainly: responding to the shrinking pool of military recruits from population decline.Source: UPI source

Weave Robotics prices a mobile humanoid robot at $7,999 for the consumer market · humanoid ⚠️ Preorder terms

Its first mobile humanoid robot, Isaac 1, comes with a full price of $7,999 and a $449/month payment plan, with first shipments set for fall 2026. The company has already shipped its home laundry-folding robot, Isaac 0, within California, and sold several Isaac robots to customers in the San Francisco area. The price genuinely pulls humanoid robots into the psychological range of premium home appliances — but it's worth noting that what's shipped so far is the task-scoped Isaac 0, not the mobile humanoid Isaac 1. The two aren't in the same tier of mechanical complexity or home-safety requirements; what's being sold right now is trust in the first delivery batch, not off-the-shelf capability.Source: Zhenji Zhineng source (WeChat, CN)

Shanghai Electric unveils embodied factory robots and an "AI-native" factory concept at WAIC · industrial ⚠️ Demo terms

This industrial conglomerate, known for heavy equipment and power infrastructure, unveiled a full lineup of embodied-intelligence robots for production lines at WAIC 2026, along with an "AI-native" smart factory model — not layering AI onto legacy equipment, but having planning tools, robots, machine tools, and control rooms share the same data stream from day one. The positioning is that of an integrator capable of delivering robots, software platforms, and monitoring systems as a bundle. What's conspicuously absent from the public disclosure is exactly what buyers would ask first: pricing, a commercial rollout timeline, and the number of pilots currently running.Source: Mobilités Urbaines source

IV. Industry Developments

Waymo notifies Uber: self-owned app launching in Austin and Atlanta in January 2028, ending exclusive partnership · autonomy

An Uber spokesperson confirmed to CNBC that the company has received notice from Waymo that it will launch its own app in Austin and Atlanta in January 2028, running alongside the existing Uber channel. The two companies' contracts covering these markets run at least through May 2028, during which hundreds of Waymo robotaxis will remain on the Uber platform — and Uber, in turn, gains room to bring in other autonomous vehicle fleets in both cities. This isn't an isolated move: the two companies already ended their partnership in Phoenix earlier this year, Waymo now operates independently in 9 other markets, and it ran a non-exclusive partnership with Lyft in Nashville last year. The Financial Times reports that Waymo has internally discussed a complete split from Uber, with friction points including conflicting regulatory proposals the two companies have pushed in different states; Uber's CTO published a video earlier this year publicly criticizing Waymo vehicle behavior as "dangerous and unsettling," and in May, Uber's CEO questioned — without naming names — how autonomous vehicles perform around school zones and in emergencies on an earnings call. Uber's stock fell more than 4% the day the news broke (per CNBC; other financial outlets reported a drop of about 9% that day — the figures diverge). The fight over robotaxi distribution rights is shifting from "whose vehicles are safer" to "who controls the passenger gateway."Source: CNBC source · TechCrunch source

Hyundai Motor Group announces transformation into a "Physical AI Solutions Company," co-building a robotics reference platform with Nvidia · autonomy

Hyundai Motor Group Executive Chair Chung Euisun laid out the group's Physical AI roadmap at an AI summit in San Francisco, attended by around 150 South Korean corporate executives, U.S. tech company executives, entrepreneurs, and students. In his words, the group is "moving beyond the boundaries of traditional automaking, expanding into autonomous driving, robotics, and AI-defined factories, and accelerating its transformation into a Physical AI Solutions Company." The vision progresses across three layers: smart devices such as vehicles and robots → smart spaces such as AI factories → and ultimately city-level integrated intelligence. Hyundai's stated point of differentiation is a data flywheel — using large-scale manufacturing, mobility, robotics, and service operations to continuously feed real-world industrial operating data back into its AI. On the partnership front, beyond co-developing a "Robot Reference Platform" with NVIDIA, the group also cited its partnership with Waymo and Boston Dynamics' strategic collaboration with Google DeepMind, and said it will invest in local ecosystem projects such as the Saemangeum AI Valley.Source: PR Newswire source

Dexterous hands remain the real bottleneck for mass production: hands lifting 5 kg break after about 50 cycles, top products last only 300,000-500,000 cycles · hardware

A robotics company executive who has already shipped to partners (using the pseudonym Wang Hao) told Yicai that the Yinshi Robotics dexterous hand his company procured broke after just about 50 cycles of handling objects around 5 kg in a warehousing and logistics scenario; given that these hands cost under RMB 10,000, he considers this "understandable" — but the bar for industrial use is tens of thousands of cycles without failure, since once a robot is deployed in a factory, hands can't be swapped out every time one fails. Yinshi responded that its load-bearing tests have reached up to one million cycles, that Wang Hao's case was an isolated incident "not particularly representative," and disclosed that the company shipped over 10,000 units of dexterous hands last year to around 700 customers⚠️ conflicting claims. A harder data point comes from Geek+ (Chinese warehouse robotics company): Dong Qinpeng, head of its embodied intelligence R&D, says warehouse robots need to grasp 8,000 to 10,000 times per day, while today's leading dexterous hands on the market last only 300,000 to 500,000 cycles — and that's still a low-degree-of-freedom product — meaning "it might break down after about a month of use." A compromise solution has already emerged: Geek+'s recently launched robot switched to a three-finger dexterous hand, covering the massive SKU variety of warehousing while leaving room for reliability and lifespan design. The gap between the elegant five-finger hands on display floors and the practical two-finger grippers on production lines is precisely the cost of a supply chain that hasn't yet been battle-tested.Source: Yicai source (WeChat, CN)

Nvidia draws Japanese industrial giants including Toyota and FANUC into the physical AI camp · hardware

Jensen Huang met this week with leaders of Japanese industrial companies including Toyota, Fujitsu, Kawasaki Heavy Industries, FANUC, and Kioxia to discuss deploying physical AI in their factories, saying "the next frontier of AI is the physical world, and this is a once-in-a-century opportunity for Japan." Kawasaki, FANUC, and Yaskawa — three robotics and automation manufacturers — already use Nvidia technology. The scale comparison is worth noting: under Nvidia's new segment reporting, data center revenue reached $75 billion last quarter, up 92% year over year, while the newly created edge computing segment (robotics, automotive, PCs) brought in only $6.4 billion, up 29%. Management expects Blackwell/Rubin chips to generate $1 trillion in revenue between 2025 and 2027 — physical AI can't yet carry the stock price on its own in the near term, but it's the next stop in this full-stack (DGX + Jetson + Cosmos) strategy.Source: The Motley Fool · Yahoo Finance source

New York school district halts "robot teacher" program after state education department pressure · adjacent

The Salamanca school district in western New York State announced Friday it is pausing its pilot bringing a humanoid robot from Realbotix into classrooms, citing the need to finalize student data privacy agreements with the state education department and to communicate further with the community. The pause came hours after State Education Commissioner Betty Rosa sent a letter stating "the department remains concerned about the potential impact of humanoid robots within high school classrooms." The project cost $57,590, was approved by the local school board last month, and had planned to introduce the robot — named Sally — this fall: it features silicone skin, can perform upper-body movements and facial expressions while remaining seated, and comes paired with an AI teaching-assistant avatar students can access on their laptops. Realbotix's CEO said the company respects the school's decision. Public acceptance of humanoid form factors in public-service settings is emerging as a constraint independent of technological maturity.Source: New York Focus source · The Guardian source

BYD officially enters humanoid robotics, with in-store sales assistance as its first deployment scenario · humanoid ⚠️ Statement only

BYD has confirmed it is developing humanoid robots in-house, with the first deployment scenario being in-store smart sales assistance — handling standardized services like product explanations and customer reception. This echoes an earlier statement from the company's Executive Vice President Li Ke, who said the robots could eventually be sold to consumers at 4S dealerships once the technology matures. There is currently no verifiable disclosure of the product's form factor, mass-production timeline, or deployment scale; a previously circulated figure on specific deployment numbers was denied by the company as inaccurate.Source: Autohome source

Hikvision Robotics reports H1 revenue of RMB 4.028 billion · industrial

The machine vision and mobile robotics business reported first-half revenue of RMB 4.028 billion. Amid the dense flow of humanoid-robotics narratives, the actual revenue scale of industrial machine-vision/AMR companies like this one serves as a yardstick for how far "embodied intelligence" still is from sustainable cash flow.Source: EEFocus source

Hardware · Supply Chain

· Dexterous hand demand: per GGII data, dexterous hand sales in China are projected to surge from 19,200 units in 2025 to 70,200 units in 2026, an increase of over 260%; the resulting downstream volume is pushing capital into segments like micro coreless-cup motors and ultra-thin reducers previously dominated by European, U.S., and Japanese manufacturers source (WeChat, CN)

· Planetary roller screws: research reports state that a single humanoid robot requires 14 planetary roller screws (for linear joints) and 12 miniature ball screws (for dexterous hands), with planetary roller screws accounting for 19% of value — the single highest-value core component in the whole machine⚠️ research-report claim source (WeChat, CN)

· Aoyi Technology: dexterous hand orders surpassed 10,000 units in H1 (previously reported); newly disclosed that its next-generation product will launch in August, still using a linkage design with 16 degrees of freedom, striking a balance between heat dissipation and transmission efficiency source (WeChat, CN)

· Axera: announced the formation of a wholly owned subsidiary, "Axera Compute," extending further into the compute side source

FutureX · Physical AI Daily — Issue 68 (07/25)

Shawn — Fri, 24 Jul 2026 14:43:26 +0000

Today's Highlights

· AgiBot (Zhiyuan Innovation) officially launched its Hong Kong IPO process, shipping 5,100 units in 2025 — the world's top general-purpose robotics maker by volume

· Galbot's T800 rolled off the line at its Zhengzhou base for the first batch, mass production shipping worldwide

· Chinese-developed UniWorld-View topped the world model leaderboard run by Fei-Fei Li's team, with code and weights fully open-sourced

· Wang Xingxing (Unitree founder) appeared on the cover of TIME with the manned mech GD01, and Unitree unveiled its wheeled-legged robot As2-W

· XPeng began small-batch trial production of its humanoid robot at its Guangzhou plant, targeting in-store sales roles by 2027

I. Paper Progress

GS-Agent: Multi-Agent + Physics Engine in the Loop for One-Sentence Controllable 4D Physical World Generation · world-model

The task of "building a 4D world" is decomposed into collaboration among multiple specialized agents, with a physics engine kept in the loop, enabling end-to-end generation of physically plausible, controllable dynamic 4D worlds from natural language — interactions among liquids, soft bodies, and rigid bodies, plus cinematic-grade camera lighting, can all be adjusted. Compared to prior generative methods that often fail on physical plausibility and controllability, this offers a new paradigm for "physical AI training grounds + content generation." Team from MIT/UMass.

Hongxin Zhang et al. (UMass Embodied AGI, etc.) · arXiv 2607.21522 source

Robostral Navigate: An 8B Monocular-RGB-Only Navigation Model Setting a New Continuous-Environment SOTA · autonomy

A navigation "recipe" designed for large-scale deployment: using only monocular RGB, an 8B vision-language model directly "points" to the next waypoint within the current frame; operating purely in image space lets the policy naturally generalize across wheeled/legged/drone embodiments without recalibration. Training on 2.4 million simulated trajectories with prefix caching compresses training tokens by 22x, shrinking the training cycle from months to days. On R2R-CE, it achieves a 77.4% success rate with a single camera — 10.5 percentage points above the previous best monocular method, and also surpassing the strongest systems that rely on depth or multiple cameras.

Arjun Majumdar et al. · arXiv 2607.20785 source · HF↑7

PhysCoRe: Adding a "Physics-Correction Residual" Layer to World Models for Deformable Objects · world-model

Predicting how deformable objects evolve under manipulation is a long-standing challenge — per-object fitting of material parameters is slow and doesn't generalize, while pure learning-based methods often fail to extrapolate and violate basic physics. PhysCoRe learns only a residual correction on top of a physical structural skeleton, balancing material-aware generalization with physical consistency.

Haocheng Yin et al. · arXiv 2607.20653 source

Reusable Skills "Emerge" in MoE VLA Models · vla

Without presetting any task decomposition or hierarchy, can a VLA model — simply equipped with a simplified Mixture-of-Experts action head — spontaneously "emerge" with the ability to break tasks down into reusable, interpretable primitives? The authors present affirmative evidence, offering a new explanatory path for compositional generalization in VLA models.

Shlok Shah et al. · arXiv 2607.20771 source

FELT: Generating Tactile Signals Directly from Vision, Bypassing the Tactile Data Shortage · perception

Touch is critical for manipulation, but tactile sensors are fragile and hard to standardize, leaving tactile data far scarcer than visual data. FELT uses vision to generate tactile signals for feeding visuo-tactile policies, addressing manipulation shortfalls when vision is occluded or ambiguous.

Zinan Li et al. · arXiv 2607.20683 source

TableVerse: A Large-Scale Tabletop Manipulation Dataset with Real-World Layouts · benchmark

The ceiling for general-purpose manipulation policies is constrained by large-scale, high-fidelity scene data. TableVerse provides a large-scale tabletop dataset with real-world grounded layouts, targeting the data bottleneck for "generalizable manipulation."

Boyuan Wang et al. · arXiv 2607.21017 source · HF↑5

Other papers today: HyWorldVLA enhances end-to-end autonomous-driving VLA with hybrid world modeling (arXiv 2607.20988 source); AXIS, a community-crowdsourced data engine for scalable manipulation (arXiv 2607.21588 source); GLAM-SLAM, real-time large-scale Gaussian splatting mapping (arXiv 2607.21416 source); Beyond Episodic Evaluation reveals memory-architecture bottlenecks in sequential embodied question answering (arXiv 2607.21571 source); URF, a unified robot control-policy framework for stable contact (arXiv 2607.20912 source); VoLN, a purely vision-based long-horizon navigation paradigm and benchmark (arXiv 2607.21400 source).

Open Source · Tools · Benchmarks

· UniWorld-View: an open-source unified world model from Tuxiang Intelligence (Chinese startup), which topped the world model leaderboard run by Fei-Fei Li's team; its training data came from Yuankong Intelligence, a startup founded by Peking University alumni, and it runs on Chinese Ascend compute, with code and weights fully open-sourced source (WeChat, CN)

· LingBot-World: an open-source interactive world model that progressively trains a general-purpose video generation model into a real-time world simulator controllable via keyboard actions and text instructions source (WeChat, CN)

II. Financing & Deals

Suzhou Junduo Robotics (China) | New Round | Amount Undisclosed · embodied

The round was led by Silicon Harbor Capital, Dianliang Capital, and Songshan Lake Science City Venture Capital, with participation from Midea Group, Midea Capital, Bangsheng Capital, GF Xinde, and CITIC Investment, among others. Junduo was founded by a team of Harbin Institute of Technology PhDs and focuses on robotic end-effectors such as electric grippers; the entry of industrial capital like Midea signals heating interest in the end-effector track.Source: Suzhou Businessperson source (WeChat, CN)

Yixin Zhinao (China) | Seed Round | Tens of Millions of RMB · world-model

Led by Haiyuan Capital. The team consists of postdocs from the University of Science and Technology of China and CUHK-Shenzhen, and advocates rebuilding robots' cognitive foundation with a brain-inspired architecture, betting on foundational innovation in cognitive architecture.Source: iFeng Tech source

Bit Infinite (China) | Angel+ Round | RMB 300 Million · adjacent

A Chinese industrial 3D large-model company, disclosed by IT Juzi on the same day alongside overseas industrial physical AI company Atoms (whose $1.7 billion raise was previously reported), showing continued capital inflow into "physical AI foundational capabilities" on the industrial side.Source: IT Juzi source (WeChat, CN)

Gravis Robotics (Switzerland) | Proposed SoftBank Acquisition | Valuation Possibly Over $500 Million · embodied ⚠️ Rumor-level sourcing

SoftBank is evaluating an acquisition of Gravis Robotics, an autonomous construction/machinery robotics company incubated at ETH Zurich, as it builds out its robotics footprint outside China; the deal has not yet been finalized.Source: Startup Fortune source

Dobot Robotics (China) | ChiNext IPO Approved | Aiming to Raise ~RMB 1.2 Billion · embodied

Following a co-founder's equity-dispute whistleblower complaint filed just before the July 20 review meeting, Dobot Robotics' IPO has now been approved, with proceeds primarily directed toward R&D and industrialization of multi-legged and humanoid robots.Source: OFweek Humanoid Robotics source (WeChat, CN)

DeepRobotics (China) | IPO Launched | Filing Submitted · locomotion

Quadruped robot maker DeepRobotics has launched its listing process; Zhejiang University professor Zhu Qiuguo co-founded the company as a side venture, and four of its core technical staff came from Nanjiang Robotics.Source: iFeng Finance source

III. Commercialization & Deployment

Galbot's T800: First "Henan-Made" Batch Rolls Off the Line, Zhengzhou Base Completes Mass-Production Loop · humanoid

On July 24, the first batch of full-size general-purpose humanoid robots, the T800, officially rolled off the line at Galbot's Zhengzhou Yunzhi base, and will be shipped progressively to China, North America, Europe, and Southeast Asia. The base's phase-one facility covers approximately 8,000 sq m and has completed the full process loop from incoming material inspection—assembly—final assembly—line-end testing—mass shipment; introducing automated production lines raised production efficiency by 40%. Each unit must pass 79 quality inspection items and simulations of 46 working conditions before leaving the factory. The T800 stands 1.73 meters tall and is the exclusive model for Galbot's Global Humanoid Combat League; it can be deployed in smart traffic management, commercial services, industrial, and STEM education scenarios. The Zhengzhou base will form multi-site production synergy with the Shenzhen Honghualing base.Source: Yujin Release/Henan Daily source (WeChat, CN)

Hello Inc. (Hellobike) Delivers Its First Mass-Produced Robotaxi HR1, Targeting a 10,000-Vehicle Fleet by 2027 · autonomy

The HR1 is an L4 mass-production model built by Hello Inc. for scaled operations, integrating the autonomous driving system and vehicle-grade safety redundancy at the whole-vehicle stage — marking Hello's Robotaxi shift from validation to scaled production.Source: Hello Robotaxi/Autohome source (WeChat, CN)

XPeng Begins Small-Batch Trial Production of Humanoid Robots at Its Guangzhou Plant · humanoid ⚠️ Trial-production-level sourcing

XPeng's humanoid robot has begun small-batch trial production at its Guangzhou plant, with the mass-production line in final integration testing; the company plans to place robots in stores worldwide as sales assistants by 2027, reusing its smart manufacturing and supply chain capabilities from car production. Small-batch trial production is not yet full-scale mass production, and the delivery timeline remains to be confirmed.Source: Shenzhen Commercial Daily source (WeChat, CN)

CiDi Aims to Scale Overseas Sales of Autonomous Mining Trucks · autonomy

According to Reuters, China's CiDi plans to push overseas sales of autonomous mining vehicles this year; as sales of unmanned mining-site vehicles grow, mining trucks are becoming a new frontier for Chinese autonomous driving expanding overseas.Source: Reuters source

Zhuoyu's Changzhou Plant Completed and in Production · autonomy

Zhuoyu (spun off from DJI's automotive division) has completed and begun operating its Changzhou plant; as smart driving expands into the physical world, the company is incorporating industrial manufacturing capability as a core competitive advantage.Source: Ifanr source

PepsiCo Expands Autonomous Freight Deployment with Gatik · autonomy

PepsiCo is expanding its partnership with autonomous freight company Gatik, using driverless trucks to support short- and mid-haul delivery of food and beverage products, as B2B trunk/branch-line autonomous driving continues to scale.Source: FleetOwner source

IV. Industry Developments

AgiBot Officially Launches Hong Kong IPO, Announces Rebrand, Takes the Global Lead in Shipment Volume · humanoid

On July 24, AgiBot officially confirmed it has launched its Hong Kong listing process, with its short-form company name changing from "AgiBot Robotics" to "Zhiyuan Innovation," and its legal entity renamed to Zhiyuan Innovation (Shanghai) Technology Co., Ltd. According to figures from Omdia/China Insights Consultancy, AgiBot shipped over 5,100 general-purpose humanoid robot units in 2025, making it the world's largest general-purpose AI robotics company; its 15,000th robot rolled off the line this past June. Chairman Deng Taihua stated 2025 revenue exceeded RMB 1 billion, targets several-fold growth in 2026, and aims to surpass RMB 10 billion by 2027. This follows the July 21 rumor of a "$20 billion valuation IPO push" (a valuation the company had at one point denied), with the news now confirmed; Hong Kong's stock market has already seen dozens of robotics companies lining up for listings in the short term. Having pursued an IPO just three years after founding, AgiBot is known for its multi-pronged capital strategy — an A-share controlling stake in Shangwei New Materials, the Hong Kong IPO, and subsidiary spin-offs.Source: China Securities Journal/Caixin source

Wang Xingxing Appears on TIME Cover with 2.7-Meter Manned Mech GD01; Unitree Unveils Wheeled-Legged Robot As2-W on the Same Day · humanoid

On July 23 local time, Unitree founder Wang Xingxing appeared on the cover of TIME magazine alongside the world's first mass-produced manned mech, GD01 (approximately 2.7 meters tall, priced from RMB 3.9 million), under the headline "The Big Robot Moment" — marking the first time a Chinese entrepreneur has appeared on the magazine's cover since Robin Li in 2018, an 8-year gap. On the same day, Unitree unveiled its new wheeled-legged robot, the "Chaofan Jianjiang" As2-W, claimed to reach speeds of 6 m/s, carry a 16 kg payload, and achieve over 30 km of range unloaded — positioned to compete with Boston Dynamics' Spot (starting at $74,500). Unitree also signed an agreement with Chengdu's Jinjiang district to build the country's first embodied-AI lab for basic education. The magazine cover and the mech's debut are largely image- and capability-showcasing moments; the mass-production delivery timeline still needs to be observed.Source: Kankan News/Sina Auto source (WeChat, CN)source

AMD Pushes Full-Stack into Robotics "Brains," Launches New Kria Module, Builds Robotics Partner Network to Challenge Nvidia · hardware

AMD unveiled a full-stack compute solution for robotics: a new-generation Kria module focused on real-time control and unified memory, moving robotics compute "from the body to the brain"; it simultaneously formed the AMD Robotics Partner Network (with Nota AI, MulticoreWare, and others as founding members). Multiple outlets have framed this as AMD's new challenge to Nvidia in humanoid robotics compute. Separately, it was reported that Foundation, backed by Eric Trump, has also chosen to co-develop military/industrial humanoid robots with AMD (a stated partnership intent).Source: The Robot Report/Reuters source source

Black Forest Labs Releases FLUX 3, Shifting from Image Generation Toward Video and Robotic Manipulation · world-model

Germany's Black Forest Labs released FLUX 3, which, beyond generating images and 20-second videos with audio, extends video generation models to robotic manipulation for the first time; its partner mimic robotics simultaneously launched the FLUX-mimic "video-to-action" model, now deployed on an Audi factory production line, continuing the "video generation models drive robots" world-model approach.Source: VentureBeat/Unite.AI source source

Mobileye Founder Shashua Steps Down as CEO as Company Bets Fully on Robotaxi · autonomy

Amnon Shashua, founder of Mobileye who has led the company for nearly 30 years, announced he is stepping down as CEO, with resources shifting fully toward Robotaxi and robotics at the transition — subsidiary Moovit cut about 30% of its staff, and earlier this year the company acquired Shashua's own robotics company for roughly $900 million.Source: TechCrunch/Automotive News source

Guangdong's Industrial Robot Output Grows 34.2% in H1, Humanoid Supply Chain Tops RMB 300 Billion · industrial

Official data show Guangdong's industrial robot output grew 34.2% year-over-year in the first half of the year; according to local figures, Guangdong's humanoid robotics supply chain has surpassed RMB 300 billion in scale, with places like Dongguan now able to deliver core motion metal components for humanoid robots at volume.Source: Science and Technology Daily/Greater Bay Area Voice source source (WeChat, CN)

Hardware · Supply Chain

· Lingxin Qiaoshou O30: the world's only 20-DOF dexterous hand with fully direct-drive mass production, priced at RMB 59,000 without tactile sensing and RMB 69,000 with tactile sensing; by comparison, overseas Shadow Hand (20+ DOF) costs RMB 1.6–2.6 million — a nearly 30x price gap source (WeChat, CN)

· Dexterous hand cost structure: Morgan Stanley estimates dexterous hands account for 17.3% of the BOM cost of Tesla's Optimus, and Elon Musk has said the forearm-plus-hand accounts for about 60% of the whole unit's engineering difficulty source (WeChat, CN)

· Ruiyan Zhikong: the core control of its dexterous hand is integrated on a 6×6cm six-layer PCB, with dual STM32 chips controlling 6 motors in parallel; over 90% of chips come from STMicroelectronics, using high integration to reduce failure rates and ensure batch consistency source (WeChat, CN)

· LEADERDRIVE (Green Harmonic Drive) and others have invested to found a precision robotics bearing company, as the harmonic reducer leader extends upstream into bearings and other components source

FutureX · Physical AI Daily — Issue 67 (07/24)

Shawn — Thu, 23 Jul 2026 14:52:33 +0000

Today's Highlights

· Atoms closes $1.7B in funding led by a16z, with Uber founder Kalanick betting on industrial physical AI

· After a nearly three-month freeze, China restarts robotaxi permits, with Momenta and Baidu the first approved

· Tesla's Q2 operating profit halves, down 57%, as Musk tempers expectations on Optimus mass production

· Psibot (Chinese embodied-AI startup) raises about $100M at a $1.48B valuation, becoming a new unicorn with Chery leading the round

· World models keep attracting capital: HiDream (Chinese world-model startup) closes a RMB 1.5B Series C, with three rounds in three months totaling over RMB 2.1B

1. Research Progress

HOST: A robot learns a new manipulation skill in seconds after watching one human video · manipulation

Existing imitation learning requires a time-consuming training loop for every new skill, and often causes the model to "forget" previously mastered ones. HOST (Human-to-robot One-Shot Skill AcquisiTion) acquires new skills from a single human video in seconds without forgetting existing capabilities, compressing "learn while using" from a training-time loop into a near-instant one-shot alignment — a step toward a continual-learning paradigm closer to how humans learn.

Guangyan Chen et al. · arXiv 2607.20033 source

What actually matters for whole-body humanoid motion tracking? An empirical breakdown · locomotion

Getting a humanoid to follow diverse whole-body reference motions without falling involves numerous modeling and hyperparameter choices whose individual impact is hard to isolate. This paper runs systematic ablation studies on the common factors in recent humanoid motion-imitation pipelines, and open-sources a modular framework, YAHMP, that can be trained, evaluated, and deployed on the Unitree G1 — a directly reusable engineering recipe for teams working on humanoid locomotion control.

Fabio Amadio, Enrico Mingo Hoffman (HUCEBOT) · arXiv 2607.19903 source

Extreme-RGMT: Letting humanoids learn extreme moves while keeping everyday stability · locomotion

General-purpose motion-tracking policies struggle to reliably execute rare, highly dynamic maneuvers, while specialized training often degrades already-learned everyday motions. This paper uses two-stage continual learning to resolve the generalist-versus-specialist trade-off, claiming SOTA on general whole-body motion tracking while significantly improving success rates on extreme moves.

Yubiao Ma et al. · arXiv 2607.20110 source

Toyota Research Institute: Active "drifting" on icy roads turns out to be the optimal collision-avoidance strategy · autonomy

Large-slip drifting has mostly been validated in deliberately staged scenarios, leaving its practical value uncertain. This paper from Toyota Research Institute (TRI) directly answers "when is drifting actually optimal for safety in real winter driving conditions" — on low-friction icy roads, actively entering a high-sideslip drift can, under specific conditions, be the better solution for avoiding a collision, connecting a seemingly flashy maneuver back to serious driving-safety questions.

Elliot Weiss, Thomas Lew, John Subosits et al. (Toyota Research Institute, TRI) · arXiv 2607.19484 source

Milo: The first open-source, low-cost autonomous guide-dog robot · autonomy

Real guide dogs cost roughly $50,000 to acquire and maintain, involve long waitlists, and have short service lives. Milo claims to be the first open-source, low-cost guide-robot platform, performing basic guide-dog collaborative navigation tasks — staying on route, avoiding obstacles and people — fully autonomously onboard with no need for prior environment mapping, advancing accessibility for assisted mobility.

Florian Golemo, Christopher Pal et al. (Mila and others) · arXiv 2607.19530 source

KineBench: A closed-loop checkup for "embodied world models" that bypasses inverse dynamics · benchmark

There has long been no reliable yardstick for evaluating the physical consistency of world models, and nearly all existing closed-loop evaluations rely on an inverse dynamics model (IDM) to extract actions, introducing extra error. KineBench uses explicit kinematic grounding to create the first IDM-free closed-loop benchmark, and finds that embodied video generation exhibits a "nonlinear scaling that plateaus with task complexity" — giving empirical guidance for future data scaling.

Zeyu Liu et al. · arXiv 2607.19876 source

LENS: Having a large model "simplify" cluttered scenes before handing off to the planner · manipulation

Densely cluttered multi-object stacks remain a difficult case for mainstream manipulation methods today — many objects, many collisions, hard-to-predict contact physics. LENS uses a large model to generate a scene-specific, task-specific abstraction layer on the fly, plugging in on top of existing planning/control stacks, and is shown to improve classical planning, model-based control, and a VLA model alike.

Aileen Liao, Rachel Holladay, Dinesh Jayaraman, Michael Posa (UPenn, MIT) · arXiv 2607.19633 source

Other papers today: Koopman Dreamer, using spectrally-constrained deterministic latent dynamics to stabilize long-horizon "imagination" in world models (arXiv 2607.19719 source); Diffusion ReRoll, revisable denoising that improves out-of-distribution performance for robot sequence prediction (arXiv 2607.19919 source); Defer to Plan, adaptive multi-agent fusion for end-to-end V2X driving (arXiv 2607.19774 source); SeededGrasp, using a VLM to predict "seed points" guiding grasping in complex multi-embodiment scenes (arXiv 2607.20207 source); EgoRecovery, learning error-correction ability from first-person human "failure-recovery" videos (arXiv 2607.19745 source); ReferTrack, monocular "refer-then-track" embodied visual tracking (arXiv 2607.20061 source); injecting offline supervision into large-scale VLA reinforcement learning to balance efficiency and out-of-distribution generalization (arXiv 2607.19399 source); NavVerse, an indoor-to-outdoor continuous navigation benchmark (arXiv 2607.19695 source); DINS-IO, self-supervised inertial odometry without position labels (arXiv 2607.20232 source).

Open Source · Tools · Benchmarks

· Unitree UnifoLM-OminiA-0.3: an omni-modal real-time interactive embodied large model, adapted for the mass-production benchmark G1, focused on "single model, full pipeline + direct voice control," the third publicly released node in the UnifoLM family source (WeChat, CN)

· AIRobot (Chinese industrial-VLA startup): with just 0.89B parameters, claims to top an industrial embodied-AI leaderboard and "surpass π0" ⚠️ vendor claim, positioned as a small-parameter end-to-end policy for industrial scenarios source

2. Funding and Deals

Atoms (US) | New round | ~$1.7B · industrial

Led by a16z, with Ben Horowitz joining the board; co-investors include Bain Capital, Fifth Wall, K5 Global, SV Angel, and Uber — the company its founder was forced out of. Bank of America, Goldman Sachs, and JPMorgan are listed as debt partners. Atoms was founded by Uber co-founder Travis Kalanick and focuses on industrial AI for physical sectors such as food, mining, and transportation — using software, sensing, and robotic automation to handle "manufacturing, moving, and warehousing." This is one of the largest single equity raises in physical AI this year, and marks the scale of top-tier US VC bets on industrial-side embodied AI.Source: TechCrunch source

Psibot (Chinese embodied-AI startup) | New round | ~$100M | $1.48B valuation · embodied

Led by automaker Chery, with participation from Lens Technology (a components/sensor supplier to Apple and Tesla) and others; founded in 2024, the company has raised roughly $300M to date and becomes a unicorn with this round. Psibot's focus is end-to-end reinforcement learning embodied models — the Psi R0/R0.5/R1 series — targeting long-horizon task execution, and has already run small-scale sorting validations in a logistics customer's warehouse.Source: Bloomberg, Tech in Asia source

HiDream (Chinese world-model startup) | Series C | RMB 1.5B (over RMB 2.1B across three rounds since April) · world-model

Following an over-RMB-500M round in April and a hundred-million-yuan-class round in May, the company closed another RMB 1.5B Series C in July — three rounds in three months totaling over RMB 2.1B. The funding cadence itself is an industry signal: capital is shifting focus from language models toward "world models," with perception-simulation-interaction of the real world becoming the hottest target. The same day also saw PrimalVerse, a Tsinghua-affiliated team, close a seed round of several hundred million RMB betting on a foundational 4D world model.Source: AI How source (WeChat, CN)

Ropedia (Singapore) | Pre-A | $30M · adjacent

The core team comes from Tsinghua University, Meta, Google, and UC Berkeley, backed by angel investors from Google, Nvidia, and Amazon. Ropedia builds physical AI data infrastructure — proprietary wearable/mobile capture devices that collect, reconstruct, and perceive multimodal human experience in real-world settings, claiming to cut robot-training data collection costs by roughly 50x. Data infrastructure is becoming another hotspot for embodied-AI funding.Source: SiliconANGLE source

LumiBot (Chinese dexterous-hand startup) | Angel round | Tens of millions of RMB · hardware

Led by a fund under Changyou Technology, with two state-owned platform funds co-investing. The founder is a former Wall Street investor; the company focuses on dexterous-hand manipulation solutions and touts what it calls the world's first actively-cooled dexterous hand, Lumi-Dex, saying it has begun sales, with this round's funds going toward mass-production ramp-up.Source: cyzone.com source

Horizon Robotics (9660.HK) | Planned bond issuance | $450M zero-coupon convertible notes · autonomy

To be underwritten by Goldman Sachs and Morgan Stanley, bringing cumulative fundraising to over RMB 20B. As a maker of autonomous-driving chips and solutions, this low-interest convertible bond replenishes capital reserves at a time of deepening cooperation with Volkswagen.Source: Sina Finance source

Previously reported, skipped this round: Weilai Buyuan's near-RMB-1B Pre-A, Agibot's push toward a $20B-valuation IPO, Yuequan Bionic's Pre-A+, and Shihe Robotics' Series C — no material new developments this period (already reported).

3. Commercialization and Deployment

Baidu's Apollo Go begins fully driverless testing in Hong Kong on the 27th · autonomy

In step with the permit thaw in mainland China, Apollo Go is starting fully driverless road testing in Hong Kong, set to formally launch on the 27th — a step toward real driverless operations in a market beyond mainland China. Compared with just holding a road-test permit, this marks a concrete move from "getting licensed" to "on the road."Source: Autohome source

Boston Dynamics' Spot deployed for autonomous inspection at Copper One mine · industrial

Mariana Minerals has deployed Boston Dynamics' quadruped robot Spot at the Copper One copper mine, replacing human workers entering high-risk zones for autonomous inspection — a real production deployment, not a demo. Mine inspection is one of the few scenarios where legged robots have already achieved commercial closure.Source: Mining Weekly source

Fortescue and Liebherr deploy 475 zero-emission autonomous machines at Australian mine · autonomy

Mining giant Fortescue and Liebherr plan to deploy 475 units of zero-emission autonomous machinery at an Australian mine site — a large-scale order for bulk mining automation, pointing to continued scale-up of heavy autonomous vehicles in enclosed industrial settings.Source: Electric & Hybrid Vehicle Technology International source

Renault accelerates warehouse automation, adopts Exotec Skypod sorting system · industrial

Renault has introduced the Skypod warehouse robotics system into its logistics operations, advancing vertical storage and retrieval automation for its spare-parts warehouse — another case of warehouse robotics landing on the automaker supply-chain side.Source: Eureka Magazine source

4. Industry Developments

After a nearly three-month freeze, China restarts robotaxi permits, reopening the driverless-vehicle race · autonomy

According to Bloomberg, Chinese regulators, after completing an industry-wide safety review, resumed issuing robotaxi permits in certain cities starting late June, with Momenta and Baidu among the first approved. Permits had been frozen since April 29, triggered by an incident on March 31 in Wuhan where over a hundred Apollo Go vehicles simultaneously stalled, briefly stranding passengers on a highway. The restart means Baidu, Momenta, and others return to expansion mode, and signals Beijing's intent to keep directly competing with Waymo in autonomous driving. This is the policy signal with the greatest impact on the pace of robotaxi commercialization in China this week.Source: Bloomberg source

Tesla's Q2 profit halves as Musk tempers expectations on Optimus mass production · humanoid

Tesla's Q2 revenue was $28.2B, up 26% year-over-year, with deliveries of 480,000 vehicles setting a Q2 record; but operating profit plunged 57% to about $400M, with operating margin collapsing from 4.1% to 1.4%, and free cash flow turning to a net outflow of $1.1B. As core automotive profitability deteriorates, the market is worried about whether Tesla can sustain its massive investments in Robotaxi, Optimus, and AI infrastructure. Musk called the humanoid robot "the hardest manufacturing ramp in Tesla's history," saying almost no component has an existing supply chain ⚠️ company statement; he also revised the Robotaxi rollout language for Phoenix and Las Vegas to "still in preparation." On Optimus, the company says mass production will begin this year, with the first units delivered to an internal "academy," and part of the Fremont Model S/X line has been converted to Optimus production.Source: Tesla Q2 earnings, Reuters source

Alibaba's Amap unveils "five-in-one" embodied AI stack, touting a unified brain and body · world-model

According to the South China Morning Post, Alibaba's Amap has packaged five components — including perception, mapping, and spatial understanding — into a robotics AI stack, claiming to give robots a "unified brain and body" ⚠️ vendor claim. This is another instance of a major internet company transplanting mapping/spatial capabilities into embodied AI, pointing to positioning around "spatial intelligence as infrastructure."Source: South China Morning Post source

Amazon plans 300-acre robotics manufacturing plant in the US · industrial

Amazon plans to build a 300-acre robotics manufacturing plant in the US, further internalizing warehouse/logistics automation capacity. As one of the world's largest users of warehouse robots, its in-house capacity build-out carries implications for both upstream supply chains and deployment pace.Source: Sohu source

Nvidia PhysX co-founder Zhang Lihua enters physical AI, building world model Fysiverse · world-model

According to an exclusive report, Nvidia PhysX physics-engine co-founder Zhang Lihua has moved into physical AI, leading a team building a next-generation physics-based world model, Fysiverse. A veteran of physics simulation moving into world models reflects how the "differentiable physics + generation" approach is drawing in core talent.Source: NetEase Mobile source

Hardware · Supply Chain

· Leadshine Technology: focused on core robot components, prioritizing joint modules + dexterous hands source

· Jiangsu Leili × MINDSYNC: leveraging three decades of micro-motor expertise to enter dexterous hands, targeting "fingertip stiffness" as a key end-effector pain point source

· Inspire-Robots: humanoid five-finger dexterous hand shipments topped 10,000 units globally in 2025 with over 60% market share; micro servo-cylinders now in production at Xiamen Software Park source (WeChat, CN)

5. This Week's Observations

US weighs banning Chinese humanoid robots, geopolitical rivalry extends into embodied AI · humanoid

According to the South China Morning Post, as US-China tech competition escalates, the US is evaluating import/use restrictions on Chinese humanoid robots. At the same time, Chinese officials have said they will encourage foreign investment into humanoid robotics R&D and related fields. The embodied-AI industry is rapidly being drawn from a pure technology race into supply-chain and market-access geopolitics — worth continued tracking for its impact on overseas expansion paths.Source: South China Morning Post source

MIIT forms data-standards working group; UBTECH leads eight national humanoid-robot standards · humanoid

China's Ministry of Industry and Information Technology (MIIT) formed a TC8 Data Standards Working Group this week and held its first symposium; UBTECH is leading the drafting of eight national standards for humanoid robots, including China's first national standard for evaluating humanoid-robot dataset quality, now formally initiated — signaling a shift in data standards from "existence" to "quality, trustworthiness, and usability." Internationally, the Robot Data Factory (RDF) standard proposed by SK Telecom was adopted by ITU-T for standardization — data standards emerged as a shared focus for regulators this week.Source: Sina Finance source

Washington, D.C. sets clear rules for robotaxis: licensing deposit + per-mile mileage tax · autonomy

Washington, D.C. plans to require a seven-figure-dollar-level deposit for robotaxi operating permits and impose a mileage tax of $0.15 per mile, aimed at preventing empty vehicles from clogging streets and internalizing external costs. This is a concrete example of Western cities shifting from "permitting the technology" to "constraining it with rules," a real cost item for driverless-vehicle unit economics and expansion pace.Source: Hola Chuhai source (WeChat, CN)

Report data: diverging figures on H1 embodied-AI funding in China, unicorn count surges · embodied

According to IT Juzi, total embodied-AI funding in China in H1 2026 exceeded RMB 93.5B across 322 deals, with 26 unicorns valued over RMB 10B (another source cites RMB 104.1B — the figures use different methodologies and scopes and shouldn't be directly compared). GGII released its "H1 2026 China Dexterous Hand Industry Survey Report" over the same period, identifying dexterous hands as the "last inch" of embodied-AI deployment and the segment with the fastest funding and growth; a separate deep-dive report on servo motors estimates demand of 1.2 million sets of joint motors, a market of RMB 91.7B, and a domestic-production rate of 80%. On the cautionary side, the PwC AI Institute warned that "blindly betting on humanoid robots is irrational."Source: AI Intelligence Cube source (WeChat, CN), GGII Humanoid Robotics source (WeChat, CN)

This week's supply-chain roundup: planetary roller screws are the tightest bottleneck, chip localization accelerates · hardware

This week's supply-chain discussion centered heavily on planetary roller screws. These are the core transmission component for humanoid linear joints — a single Optimus unit uses over 30 of them; the bottleneck lies not in design but in manufacturing — constrained by grinding-machine lead times and a shortage of skilled grinding operators, the industry widely reports capacity ramp-up being stuck (one estimate pushes a leading maker's "1,000 units/week" target from July to September), with the domestic-production rate still under 15%. For reference, the global planetary roller screw market was worth only about $1.27B in 2022 — meaning a single customer's annual demand could rival the entire prior global market. A loosening signal: China's first domestic mass-production line came online this week, with Chinese-made units priced at roughly 40-50% of comparable overseas products — supply is being pulled up, but high-end capacity and consistency remain real problems.
A second thread is domestic production of robot-specific chips: at WAIC, several leading players (Agibot, UBTECH, TRON1 and others) have already put Chinese-made MCU/control chips into their robots, with domestic chips moving from "backup" to "standard," and a single-chip full-stack architecture seen as the watershed for scaled deployment (Sunrise's Digua S600 has already locked in 20+ clients); on the cost side, a single humanoid robot often uses 100+ MCUs, with chip BOM cost reaching $3,000-5,000 — a main battleground for domestic substitution. On the battery side, CATL brought online the world's first humanoid-robot battery PACK production line this week, directly tied to humanoid mass production and cost.Source: Neural Leap source (WeChat, CN), AI Melon Field source (WeChat, CN), Humanoid Robot Full-Stack Lobster source (WeChat, CN)

FutureX · Physical AI Daily — Issue 66 (07/23)

Shawn — Wed, 22 Jul 2026 14:58:56 +0000

Today's Highlights

· Johnson & Johnson's Ottava surgical robot receives FDA approval, as a medical-device giant formally enters laparoscopic robotics

· Home robotics startup Futuring Robot closes its third funding round in under a year, raising nearly RMB 1 billion in its Pre-A round, becoming the sector's first unicorn

· Fei-Fei Li's World Labs acquires robotics simulation company SceniX, turning generated worlds into training grounds

· US-listed surgical robotics company Vicarious Surgical votes for bankruptcy liquidation, having raised over $2 billion cumulatively

· Fei-Fei Li's team unveils "Masked Visual Actions": a unified world model trained on just 15 hours of data

I. Research Progress

Masked Visual Actions: painting "actions" into pixels, training a unified world model on 15 hours of data · world-model

Video models naturally learn how the world moves, makes contact, and responds — the hard part has always been feeding a robot's "actions" into them in a form they can understand. Stanford's Fei-Fei Li team proposes Masked Visual Actions: representing actions as the "partially revealed trajectory" of some entity in a video — reveal the robot's motion, and the model performs forward dynamics prediction of the scene's response; reveal the target object's desired motion, and the same model can infer the robot action that produces that outcome. Fine-tuned on just 15 hours of real and simulated masked samples, a single set of weights performs policy evaluation, model-based planning, and inverse action synthesis across multiple scenes and embodiments.

Hadi Alzayer, Wenlong Huang et al.（Stanford）· arXiv 2607.19343 source · HF↑3 · Coverage: PaperWeekly source (WeChat, CN)

ergoCub: building "human burden" into hardware optimization, published in Nature Machine Intelligence · locomotion

A team at the Italian Institute of Technology (IIT) has published the humanoid robot ergoCub in Nature Machine Intelligence. Building on iCub3, rather than treating hardware as a fixed chassis, the team uses coupled human-robot dynamics to jointly optimize "ergonomic burden during load-carrying" and "the robot's own locomotion," then translates the optimized structural proportions into a real, fully assembled robot — turning human burden into an optimizable design variable for scenarios like collaborative carrying that reduces worker load.

IIT (Italian Institute of Technology) · Nature Machine Intelligence · Coverage: 科研灵感站 source (WeChat, CN)

Agentic Real2Sim: letting VLM agents automatically bring real-world scenes into runnable simulation · world-model

Real-to-sim (real2sim) has long relied heavily on manual work, since it requires not just reconstructing geometry but also recovering object states, inferring physical parameters, and assembling camera, pose, and trajectory data into a runnable physics simulation. This work automates that pipeline using vision-language agents, pushing toward "generate trainable simulation from a segment of real-world interaction" and lowering the cost of collecting embodied data.

Guanxiong Chen et al. · arXiv 2607.19190 source

WorldScape Policy 2.0: adding "reasoning-augmented memory" to world-action models · world-model

World-action models (WAMs) perform manipulation by jointly modeling visual state transitions and robot actions, but tend to "forget" over long-horizon tasks. Version 2.0 introduces a reasoning-augmented memory mechanism, making world modeling more controllable and steerable over long time horizons and improving consistency across multi-step manipulation.

Haisheng Su et al. · arXiv 2607.18840 source

RoboInter1.5: a full suite of "intermediate representations" for embodied world modeling · benchmark

Existing robotics datasets are expensive, tied to a single embodiment, and lack the fine-grained annotations needed for generalizable reasoning and long-horizon environment dynamics. RoboInter1.5 provides a holistic suite of intermediate representations for embodied world modeling and manipulation, aiming to fill the gap in "reusable, cross-embodiment" structured annotation as a general-purpose interface between world models and policies.

Ziqin Wang et al. · arXiv 2607.18709 source

Athena-Brain: a technical report on an efficient "robot brain" for edge deployment · vla

As embodied agents grow more capable, the industry needs a compact "brain" that can fit onto a robot's own hardware while retaining large models' general knowledge and reasoning. Athena-Brain optimizes for efficiency in edge deployment, aiming to balance general intelligence with high-level interaction with the embodied environment on-device.

Jialian Li et al. · arXiv 2607.18985 source

Other papers today: STeP uses signal temporal logic to give VLAs precise spatial/temporal/logical constraint expression (arXiv 2607.18580 source); MVP-Tac, a miniature visuo-photoelastic dual-mode tactile sensor for minimally invasive surgery (arXiv 2607.18660 source); The Open Ant, an open-source low-cost real-robot reinforcement learning platform simplifying sim-to-real (arXiv 2607.18488 source); No Training, Better Flights uses test-time scaling to improve VLM drone navigation without training (arXiv 2607.19288 source); Recti-Q, out-of-distribution-robust quantization-aware inference for edge robots (arXiv 2607.18540 source); MAGE uses multimodal agents for chip macro-cell placement (arXiv 2607.18536 source).

Open Source · Tools · Benchmarks

· NVIDIA open-sources a medical physics simulation framework: Nvidia released what it calls the "first" GPU-accelerated medical physics simulation framework for training and validating surgical robots, and demonstrated predictive surgical robot capabilities live with CMR Surgical source

· DM0.5: an open-source embodied model that took first place on a simulated manipulation benchmark with a real-robot score of 54.42, advancing alignment between open-source embodied models and real-robot success rates source

II. Funding & Deals

Futuring Robot｜Pre-A round｜nearly RMB 1 billion cumulative · embodied

This round saw continued follow-on investment from existing shareholders Yuanlai Capital, Linkedu Capital, and Boyu Capital, joined by a leading Chinese internet company, CHJ Capital (Inovance's investment arm), Guofang Venture Capital, and Deling Capital, among others. This marks the company's third funding round this year, bringing its cumulative total to nearly RMB 1 billion — a record for the home general-purpose robotics sector. The company focuses on consumer-facing general-purpose home robots and says its products have entered 500+ households. Amid heavy capital inflows into dexterous hands and humanoid platforms, this rare bet on a consumer-facing home robotics company is a footnote to the current shift in robotics funding: renewed enthusiasm, but with sharper focus on orders and real-world deployment.Source: 极客公园 (GeekPark) source (WeChat, CN)

Gritt (United States)｜pre-seed + Series A｜$32.4 million (≈RMB 220 million) · industrial

Of this, a $26 million Series A was led by Obvious Ventures with participation from Union Square Ventures and others. The company was co-founded by Puneet Puri, whose background includes Carnegie Mellon's Robotics Institute, and builds "physical AI" for construction sites: robotic arms plus AI, mounted as attachments onto existing job-site equipment, working alongside construction workers on high-intensity repetitive tasks like material handling and assembly. The company says it has autonomously installed tens of thousands of solar panels with zero damage and has entered job sites for top-ten US construction firms. Its entry point is the labor shortage projected as more than 41% of the US construction workforce is expected to retire by 2031.Source: 智东西 (AI/robotics media) source

Wenshi Robotics｜Series A+｜amount undisclosed · industrial

A provider of full-scenario logistics robotics solutions, this round was backed by Jinfu Capital. Following its Pre-A round in 2022 and Series A in 2023, this marks three consecutive institutional funding rounds, pointing to sustained capital inflow into the AI smart warehousing sector.Source: 第一机器人 (robotics media) source (WeChat, CN)

III. Commercialization & Deployment

Johnson & Johnson's Ottava surgical robot receives FDA approval · adjacent

Johnson & Johnson's Ottava soft-tissue surgical robotic system has received approval from the US FDA. This means a medical-device giant has formally earned entry into the laparoscopic surgical robotics market, long dominated by Intuitive Surgical's da Vinci system — a genuine regulatory milestone rather than a demo or plan, and a substantive signal for the competitive landscape in surgical robotics.Source: MassDevice source

Aurora launches second-generation driverless truck in the US · autonomy

Autonomous freight company Aurora announced the launch of its second-generation driverless truck in the US to meet customer demand. Aurora has already achieved commercial driverless heavy-truck freight operations in Texas; this second-generation launch is an expansion on top of existing real-world operations, not a first-time road debut.Source: Yahoo Finance source

Xinjiang deploys 108-arm cotton topping robot · adjacent

Xinjiang has deployed a 108-arm robot for cotton topping (removing the terminal bud), replacing significant manual labor and reducing labor costs. Open, unstructured agricultural environments have long resisted automation, and this multi-arm, purpose-built form factor is a pragmatic way to push robotics into real-world fields.Source: Big News Network source

Wetour Robotics lands $500,000 warehouse deployment order · industrial

Wetour Robotics has signed a warehouse deployment order worth roughly $500,000, and disclosed that its framework could potentially scale to as many as 20 warehouses across the US. The "20 warehouses" figure is an expansion target, not an already-delivered scale — actual deployment currently stands at this single order.Source: Investing.com

IV. Industry Developments

Fei-Fei Li's World Labs acquires robotics simulation company SceniX · world-model

World-model company World Labs, founded by Fei-Fei Li, has acquired robotics simulation company SceniX (reportedly founded by one of her students). The move directly connects World Labs' generative world models to robotics simulation and training — "using generated worlds as robot training grounds" is a substantive step in world models moving from content generation toward supplying embodied data, echoing this week's broader theme of using video/generated worlds to work around the high cost of real-robot data collection.Source: 具身纪元 (embodied AI media) source (WeChat, CN)

Surgical robotics company Vicarious Surgical votes for bankruptcy liquidation · adjacent

According to foreign media reports, investors in US-listed surgical robotics company Vicarious Surgical (NYSE: RBOT) voted to immediately cease operations and enter bankruptcy liquidation. The company had raised over $2 billion cumulatively. Amid broadly warming robotics funding, a publicly listed surgical robotics company reaching liquidation is a real-world correction to the "story without orders" playbook.Source: 机器人前瞻 (robotics media) source (WeChat, CN)

Volkswagen deepens partnership with Horizon Robotics, L3 autonomous driving to ship as early as H2 2027 · autonomy

Volkswagen Group and Horizon Robotics (Chinese ADAS chipmaker; via joint venture CARIZON) announced a deepened partnership on AI and autonomous driving, with a target of deploying L3-level capability in China as early as the second half of 2027, advancing toward L4 thereafter. This is a further step in binding a Chinese autonomous-driving solution into a multinational automaker's mass-production roadmap.Source: Automotive News source

JD.com acquires Shanghai Zhangjiang land parcel to build 105,000-sqm robotics industrial base · industrial

JD Property has acquired a land parcel in Shanghai's Zhangjiang district, where it will build a robotics-focused industrial base of approximately 105,000 square meters. Combined with recent moves by JD.com, Alibaba, and other major tech companies rushing into robotics infrastructure, this points to large Chinese tech firms shifting resources from pure large-model development toward embodied AI production lines and industrial infrastructure.Source: 观点网 (real estate/industry media) source

Kunlun Tech unveils three world models targeting embodied AI · world-model

Kunlun Tech (Chinese gaming/AI company) unveiled three world models (including the Riemann series) at WAIC, positioned to connect virtual and physical reality and serve embodied AI. The company views world models as a pathway toward larger-scale embodied AI applications; this has been interpreted externally as an "early bet placed ahead of time," with the commercialization path still unproven.Source: 每日经济新闻 (National Business Daily) source (WeChat, CN)

Anthropic reportedly held acquisition talks with embodied-model company Physical Intelligence · embodied ⚠️ unconfirmed report

Multiple foreign media outlets reported that AI company Anthropic held talks to acquire embodied foundation-model company Physical Intelligence (valued at approximately $11 billion), which Physical Intelligence subsequently publicly denied. The report itself reflects intensifying competition among leading digital AI companies for robotic "bodies," though the deal remains unconfirmed.Source: Reuters and others source

Unitree's Wang Xingxing: embodied AI's "ChatGPT moment" could arrive within two to three years · humanoid ⚠️ statement, not milestone

Wang Xingxing, founder of Unitree (Chinese humanoid/quadruped robot maker), said in a public speech that he expects embodied AI's "ChatGPT moment" — a robot completing roughly 80% of tasks in roughly 80% of unfamiliar scenarios — could arrive within two to three years at the earliest. This is an industry outlook rather than a deployment milestone, noted briefly here.Source: Xinhua Finance source (WeChat, CN)

Hardware & Supply Chain

· Solid-state / robot batteries: Reports indicate CATL has brought online what it says is the world's first humanoid robot battery PACK production line, while EVE Energy is also advancing mass-production plans for its 1.0 all-solid-state battery — upstream suppliers are building out dedicated "robot battery" capacity as a standalone product category source (WeChat, CN)

· Servo / joint motors: Industry research estimates demand for humanoid robot joint motors at 1.2 million units, a market size of roughly RMB 91.7 billion, with domestic (Chinese) supply already at roughly 80% — servo motors, as the "core muscle," are becoming a segment of high certainty for Chinese supply chains source (WeChat, CN)

FutureX · Physical AI Daily — Issue 65 (07/22)

Shawn — Tue, 21 Jul 2026 14:48:42 +0000

Today's Highlights

· Samsung establishes robotics division RX reporting directly to the CEO, poaches former Boston Dynamics strategy lead

· AGIBOT (Chinese humanoid robotics startup) reportedly pushing for IPO at a $20 billion valuation, with revenue projected at RMB 4 billion this year

· Nvidia open-sources Cosmos 3 Edge: a 4B-parameter world model that generates actions directly on-device on robots

· Momenta secures Robotaxi road-test permit in Shenzhen, stock up over 10% in a single day

· Alibaba releases 122B embodied foundation model RynnBrain 1.1, beating closed-source models across three spatial reasoning leaderboards

I. Research Progress

RynnBrain 1.1: 122B embodied foundation model beats closed-source models across three spatial reasoning leaderboards · vla

Alibaba DAMO Academy has pushed its open-source embodied foundation model to 122B-A10B scale, currently the largest open-weight tier publicly available, with HF trending at 30↑. Compared to version 1.0, the full lineup adds contact-point prediction, and the 2B/9B variants gain native 3D grounding, making outputs more tightly aligned with actual mechanical manipulation. The accompanying RynnBrain-VLA uses a unified cross-embodiment action space plus embodiment-specific masks, and has already been deployed on three robot platforms: Unitree G1, Astribot-S1, and Tianji Wuji. The 122B-A10B model surpasses all closed- and open-source models evaluated in the paper on VSI-Bench, MMSI, and RefSpatial-Bench; in real-robot experiments, policies initialized with it also outperform the Qwen series and mainstream general-purpose VLAs.

Kehan Li et al. (Alibaba DAMO Academy) · arXiv 2607.17977 source

Qwen-VLA: Manipulation, navigation, and human first-person data folded into a single policy · vla

The Qwen team is targeting a fragmentation problem: a dual-arm robot that can cook can't walk, and vice versa—manipulation arms predict joint poses while navigation robots predict waypoints, and the two have historically been hard to share in one model. Their approach uses Qwen3.5-4B as the cognitive backbone connected to a DiT flow-matching action decoder, then writes robot morphology and control mode into the input text via "embodiment-aware prompts," sharing the same latent space without changing the architecture. As a single generalist, it achieves 97.9% on LIBERO, a 76.9% out-of-distribution generalization success rate on a real dual-arm ALOHA platform, and 56.7% on RoboCasa-GR1, surpassing models specifically fine-tuned for that platform.

Qwen Team · arXiv 2605.30280 source · Analysis: Embodied Intelligence and Spatial Perception source (WeChat, CN)

Patch Policy: Dense visual features without the VLM baggage · manipulation

This work, with Yann LeCun as a contributor, points to an awkward tradeoff: robot policies either compress each frame into a single global token—losing spatial detail—or shoulder the full inference cost of a billion-parameter VLM. The authors simply add a block-causal attention mask, letting an ordinary transformer policy directly consume dense patch tokens from a pretrained ViT while preserving temporal causality. Across four simulation suites and three real-robot environments, this improves on global pooled representations by 40%; it outperforms fine-tuned OpenVLA-OFT by 18% while using only about 0.7% of its parameters.

Gaoyue Zhou et al. (NYU) · arXiv 2607.18236 source

POT-VLA: Giving humanoids memory that "the object is still there," success rate up from 39/80 to 71/80 · locomotion

A hidden failure mode in long-horizon humanoid loco-manipulation is "object state divergence": the object state used to drive whole-body actions and the state used to judge whether an action achieved its goal are actually two separate representations. The authors use persistent object tokens to maintain role-indexed 3D object records from RGB-D input, with the same record both conditioning action generation and supporting geometric predicate verification, closing the execution loop. Across eight real-world task categories on Unitree G1, the aligned GR00T-N1.7 baseline improved from 39/80 to 71/80.

Peng Ren et al. · arXiv 2607.18016 source

PolicyTrim: No retraining needed, just teaches the VLA to stop taking detours—up to 5.83x faster end-to-end · vla

What actually slows down VLA deployment is often not single-inference latency, but how many replanning cycles and real physical steps the policy needs. A Sichuan University team uses two-stage reinforcement learning post-training: first letting the model discover reliable action-chunk execution boundaries, then using a step-saving reward to suppress correction and retreat actions. On LIBERO, π0.5 achieves a 5.83x end-to-end speedup while maintaining over 98% success rate, with physical step count reduced by up to 51.4%; on a real Agilex Piper robot, the average speedup is 1.86x.

Sichuan University, Lei Yinjie's team · arXiv 2606.22540 source · Analysis: AI Era (Chinese AI media outlet) source (WeChat, CN)

Test-time scaling for world-action models: Zero-shot trajectory selection via geometric consistency · world-model

Robots can't compute an answer first and then choose, the way language models do—actions must execute immediately, so the model must judge on the fly whether spending extra compute is worthwhile. The authors exploit the fact that world-action models emit both action chunks and predicted future observations at each rollout, using a frozen geometric foundation model to compute cross-view depth reprojection consistency to rank candidates, requiring no training or task labels. RoboCasa's group average score rises from 66.3% to 68.4%; with consistency gating added, extra sampling is triggered at only 26.2% of decision points, recovering 74.8% of the gains achieved by the always-on mode.

Zesen Zhao et al. · arXiv 2607.17454 source

Adding a depth prior to JEPA world models lets an 18M-parameter model handle in-the-wild data · world-model

Complex, unpredictable outdoor real-world robot video has long been a challenge for JEPA-style world models. The authors introduce depth supervision plus isotropic latent-space regularization during training, over-parameterizing only at training time with no added inference cost. Training an 18M-parameter model on agricultural robot video, the frozen representation's error on a visual odometry probe is 33% lower than the LeWM baseline, and its "surprise" discrimination on out-of-domain TartanGround also clearly widens, with the advantage growing with rollout length.

Usman M. Khan · arXiv 2607.16314 source

"Reason before acting" isn't necessarily more stable—one class of VLA actually collapses fastest · vla

Intuitively, policies with reasoning steps should better absorb disturbances. The authors compare reasoning-free, text chain-of-thought, and latent iterative loop models head-to-head on LIBERO and SimplerEnv, and find the opposite: the latent iterative class collapses outright under random noise and white-box attacks, while the other two hold up, and adjusting reasoning depth barely helps recover it. Even more relevant for practitioners: a seemingly near-perfect "plan-action consistency" probe drops to random-chance level under adaptive attacks.

Tuan Duong Trinh et al. · arXiv 2607.17786 source

MEVION: A $14,000 open-source dual-arm data collection rig that fills ALOHA's gap in handling heavy loads · benchmark

ALOHA is already the de facto standard for dual-arm data collection, but its force and speed ceilings are low—heavy objects and fast motions are both off the table. The authors use sheet-metal welding to build a large-frame body from a small number of parts, all sourceable off-the-shelf; the four 6-DOF arms weigh 7.0 kg each with a peak torque of 60 N·m, and the full setup costs around $14,000. The elbow joint borrows a closed-chain mechanism from quadruped robots to reduce distal mass. Both hardware and software are fully open-sourced.

Kento Kawaharazuka et al. · arXiv 2607.17970 source

Other papers today: BoxTwin learns the full dynamics of elasto-plastic articulated objects from video, pushing digital twins toward deformable objects (arXiv 2607.17132 source); PhyAgentOS provides a unified runtime for heterogeneous embodied agents, separating "execution termination" from "semantic completion" (arXiv 2607.16636 source); World Translation uses inverse dynamics extraction plus unpaired domain translation to compress the sim2real gap (arXiv 2607.18154 source); FM-VLA adds force-sense memory to VLA for contact-rich manipulation (arXiv 2607.18231 source); Foresight Residual RL uses offline-estimated foresight value to optimize subtask handoff quality in long-horizon assembly (arXiv 2607.16506 source); an asynchronous multimodal diffusion policy handles the mismatch between each modality's perception rate and inference latency (arXiv 2607.17257 source); SinD 2.0 releases a multi-city signalized intersection drone dataset with annotated semantic risk (arXiv 2607.16943 source); A2RL Vmax provides a long-distance, high-speed perception and multi-vehicle interaction dataset for autonomous racing (arXiv 2607.17813 source); GeoWorldAD applies a geometric world-action model to autonomous driving planning (arXiv 2607.17521 source); UniETP unifies observation formats and action types across different simulators for general-purpose embodied task planning (arXiv 2607.18062 source); Token-Wise Latent Streaming lets a slow reasoner stream hidden states token-by-token to a fast planner, compressing observation staleness (arXiv 2607.16806 source); importance sampling plus PCA finds rare failures in a commercial autonomous trucking stack (arXiv 2607.18106 source).

Open Source · Tools · Benchmarks

· NVIDIA Agent Toolkit × Omniverse library: At SIGGRAPH, Nvidia folded its Omniverse library into the Agent Toolkit, letting AI agents build "simulation-ready" worlds directly, alongside 21 newly announced graphics and simulation research results source

· lingbot-world-2: Claimed to be the world's first hour-scale real-time world model, now open-sourced—moving the camera generates responsive scenes in real time source (WeChat, CN)

II. Funding and Deals

AGIBOT ｜ Planned IPO ｜ Target Valuation $20 billion (approx. RMB 150 billion) · humanoid ⚠️ unconfirmed report

The company has brought on CITIC Securities as underwriter, with full-year 2026 revenue projected at RMB 4 billion—already over RMB 1 billion in Q1 alone. The revenue trajectory underpins this valuation: RMB 300,000 in 2023, over RMB 60 million in 2024, and past RMB 1 billion in 2025. Unlike peers betting on research and demonstration shipments, AGIBOT has focused on industrial manufacturing—Longqi's Nanchang production line has deployed in parallel, with Furian Precision and SAIC factory lines running continuously; cumulative units shipped topped 10,000 as of March this year. Co-founder Peng Zhihui has previously said the company's biggest current shortage isn't capital but data.Source: Robo Encyclopedia (Chinese robotics industry outlet) source (WeChat, CN) · The Standard source

GigaAI ｜ New Round ｜ $3 billion valuation ｜ Planning Hong Kong listing this year · world-model

Following three rounds—Pre-B, B1, and B2—totaling RMB 3.5 billion between March and June, founder Huang Guan has publicly stated discussions are underway for a Hong Kong listing as early as 2026, aiming to be the "first world model IPO." The technical anchor is its proprietary embodied world model GigaWorld-1, which topped Tsinghua's WorldArena benchmark in April with a score of 62.34—the only embodied world model on that leaderboard to break 60 overall. Commercially, products are already deployed at FAW's die-making operations and Alibaba Cloud, with the general-purpose robot Maker H01 targeting delivery of a thousand units this year.Source: The Standard source · Robo Encyclopedia source (WeChat, CN)

Shouxing Technology ｜ Series A2 ｜ Hundreds of millions of RMB · embodied

Led jointly by CITIC Goldstone and Geely Capital, with CITIC Securities International Capital and Guotai Junan Innovation Investment following, existing shareholders Zhongwang Capital and AGIBOT increased their stakes, and Lei Jun's Shunwei Capital oversubscribed. The company specializes in bionic faces that can blink, frown, and smile, pursuing an emotional-interaction path rather than competing in whole-body humanoid platforms. Funds will go toward expanding its bionic product line and building standardized mass-production lines.Source: Tech Capital Circle (Chinese venture funding outlet) source (WeChat, CN)

Striverobotics ｜ Series C ｜ Hundreds of millions of RMB · industrial

Led jointly by Zhejiang state-owned Qianggang Capital and global optical module leader Innolight, with Guangtou Capital, Guohai Securities, and Juntong Capital following, and existing shareholder Fosun RZ Capital increasing its stake. The company builds embodied robots for high-altitude specialized work—a niche overshadowed by the humanoid boom but with among the most rigid demand for replacing dangerous human labor.Source: Tech Capital Circle source (WeChat, CN) · 36Kr source

Sereact (Germany) ｜ Series B extended to $116 million ｜ Zalando takes strategic stake · industrial

European e-commerce giant Zalando has joined this Berlin-based warehouse robotics company's Series B as a strategic investor, with Headline leading; total funding raised now exceeds $145 million. More than 200 Sereact systems are currently running across Europe, with customers including Mercedes-Benz Trucks and BOL. The funding will primarily go toward its next-generation robot brain, Cortex 2.0, whose approach is to use a learned world model to predict physical outcomes in advance and score candidate trajectories by risk and efficiency—shifting warehouse automation from "try it and see" to "compute first, then act."Source: Retail Technology Innovation Hub source

T2 (Japan) ｜ New Round ｜ Daiwa House Ventures and Mitsui Fudosan invest · autonomy

This Japanese L4 line-haul trucking logistics company secured investment from two real estate and logistics real estate giants after completing 500 km of real-world autonomous truck testing—the identity of the investors is itself a signal: Japan's commercialization entry point for driverless trucking is being staked on logistics real estate.Source: International Business Times source source

Robopoet ｜ Strategic Investment ｜ Tuya Smart · adjacent

Following a Pre-A round of over RMB 100 million this month, IoT platform Tuya Smart has entered as a strategic investor, pointing toward mass production and distribution channels for consumer-grade embodied products.Source: PR Newswire source

Industry data: Global robotics startup funding hits $18.8 billion in H1, already surpassing all of 2025 · adjacent

A Crunchbase report shows global robotics startup funding reached $18.8 billion (approx. RMB 136 billion) in H1 2026, far exceeding all of 2025's $15 billion—a historic high. The trend is even steeper in China: by IT Juzi's count, embodied intelligence funding in China reached RMB 93.5 billion in H1, roughly 5x year-over-year, with 322 deals, up 137% year-over-year, and the number of unicorns valued above RMB 10 billion rising from 3 in 2025 to at least 26. In Hong Kong, 10 companies with robotics businesses have already listed in H1 this year with 16 more in the pipeline—yet the two that listed earliest, Huayan Robotics and Ledong Robotics, have both seen their stock prices fall by nearly half from their highs, suggesting secondary-market price recalibration has already begun.Source: Medical Device Business Review (Chinese medtech outlet) source (WeChat, CN) · Robo Encyclopedia source (WeChat, CN)

III. Commercialization and Deployment

Tesla expands Robotaxi to Tampa and Orlando; Cybercab to feature built-in Starlink antenna · autonomy

Following its expansion to Miami this month, Tesla has pushed Robotaxi service to two new Florida cities, Tampa and Orlando, ahead of its Q2 earnings report. Also announced the same day: Cybercab will integrate a Starlink V5 antenna—for a driverless fleet, remote takeover and telemetry links in cellular dead zones have long been an operational hard constraint, and satellite backhaul fills that gap. Robotaxi and Optimus progress are also expected to be the focus of Wednesday's earnings call, with markets watching whether cash burn can translate into a per-vehicle profitability model.Source: Reuters source · Electrek source

Momenta Robotaxi secures road-test permit in Shenzhen, stock up over 10% in a single day · autonomy

Shenzhen becomes Momenta Robotaxi's fourth core city in China after Shanghai, Suzhou, and Wuxi, with overseas testing already underway in Munich, Germany and Abu Dhabi, UAE. The market reaction was immediate, with the company's stock up over 10% that day. A previously reported mass-production custom vehicle project with SAIC's Xiang Dao is still on track for a 2027 reveal, with SAIC manufacturing, Momenta providing the L4 solution, and Xiang Dao operating—a division of labor that shifts Momenta from a Tier 2 solution seller toward a deep partnership with the automaker.Source: Eastmoney source · Momenta source (WeChat, CN)

BMW locates humanoid robot software center in Landshut, prioritizing data and simulation before production-line deployment · industrial

BMW has designated its Landshut, Germany components plant as the central development site for humanoid robot software, handling training data generation and preparation, virtual simulation, motion planning, and robot training—splitting responsibilities from existing production-line pilots in Leipzig and Spartanburg, which validate specific applications while Landshut builds the technical foundation. The technical approach is an open modular architecture combining fixed programmed workflows with AI models, where a VLA model translates visual information and task descriptions into concrete actions; data collection uses cameras, motion-capture suits, and data gloves. Project lead Christoph Jagoda emphasized that the evaluation covers not just technical feasibility but which processes humanoid robots can genuinely deliver economic returns for. Partners include Athenyx Robotics and Landshut University of Applied Sciences.Source: Automotive Manufacturing Solutions source

National-Local Joint Center and Huawei launch embodied intelligence training-ground demonstration site, Linglong robots already deployed at Yanfeng production line · industrial

The National-Local Co-Built Humanoid Robot Innovation Center and Huawei unveiled an "embodied intelligence training-ground demonstration site" during WAIC, targeting two bottlenecks: scarcity of high-quality data and insufficient model generalization. Huawei contributes compute, storage, networking, and its CloudRobo embodied simulation-training platform, while the innovation center contributes robot hardware, data, models, evaluation, and deployment. One concrete result: Linglong-series robots incorporating the joint research have already been deployed at Yanfeng, completing real-world data collection and autonomous work verification on the shop floor—validating the transfer path from "training-ground training" to "model iteration" to "industrial field deployment." Chief scientist Jiang Lei said the demonstration site has made breakthroughs in four areas: large-scale real-world data sets, Ascend-native VLA models, cloud-edge-device coordinated architecture, and an end-to-end toolchain.Source: Huawei China source (WeChat, CN)

Incheon Coupang warehouse fire burns for 61 hours, hundreds of lithium-battery robots on the fifth floor a hazard concern · industrial

A fire at a Coupang logistics center in Incheon, South Korea took 61 hours to fully extinguish, possibly the longest on record in the area, with flames burning through six floors; hundreds of lithium-battery-powered robots on the fifth floor were at one point flagged as an explosion risk. On-site investigation points to melted roller shutter doors and extremely high fire load. This highlights an under-discussed boundary condition as warehouse automation scales up: densely packed lithium-battery-powered robots are themselves a concentrated fire load, and existing fire-safety designs may not have been built around that premise.Source: Seoul Economic Daily source · Sohu source

RoboSense forms alliance with Abu Dhabi's Origen, pushing embodied intelligence into the Middle East · autonomy ⚠️ letter-of-intent stage

RoboSense (HKEX: 2498) signed an agreement with Abu Dhabi AI company Origen during WAIC, covering spatial intelligence, autonomous driving, industrial inspection, and smart cities, with Origen handling regional delivery, solution integration, and localization. Both sides currently describe the arrangement as "evaluating pilot deployment and localization strategy," with no specific orders or installation figures yet.Source: Zawya source

IV. Industry Developments

Samsung Electronics establishes robotics division RX reporting directly to the CEO, poaches former Boston Dynamics strategy lead · humanoid

Following last week's reports that Samsung plans to build a humanoid robot base in Gumi, Samsung Electronics has consolidated its scattered robotics efforts into a new organization: the Robotics eXperience (RX) Business Promotion Office, reporting directly to CEO Roh Tae-moon, overseeing mid-to-long-term strategy, core technology development, and commercialization, based at the Seoul R&D campus. Personnel moves are the strongest signal here—Executive Vice President Lee Dong-geun, who previously led robotics strategy at Hyundai Motor Group and Boston Dynamics, has been named head of the robotics strategy team, joined simultaneously by Seoul National University professor Kim Hyun-jin (intelligent autonomous robotic systems) and Ajou University professor Kim Eui-gyeom (manipulator and manipulation technology). Samsung has explicitly named data as the key competitive edge in robotics, and will build a Data Factory in Gumi to collect and analyze real industrial environment data, along with research centers in the US, China, and Japan. Samsung's stock rose over 5% following the announcement, and Gumi's mayor has publicly pledged full support for Samsung's 19 trillion won robotics and AI investment.Source: The Korea Times source · Chosunbiz source

Nvidia open-sources Cosmos 3 Edge: 4B-parameter world model generates actions directly on-device on robots · world-model

The Cosmos 3 family gains its smallest tier: Cosmos 3 Edge, a 4-billion-parameter model built for edge deployment—about one-sixteenth the size of Super (64B), announced at GTC Taipei on May 31, with Nano sitting at 16B. Its target scenario is specific: machines in factories, warehouses, and hospitals need to run data-center-level performance on memory-constrained bodies. The architecture is a Mixture-of-Transformers dual-tower design: an autoregressive tower handles visual and text tokens for understanding and reasoning, while a diffusion tower handles visual, audio, and action tokens for prediction, generation, and neural simulation; both towers retain their own normalization layers and MLPs while sharing multimodal attention, allowing the model to reason about a scene before generating output. The reasoner uses a 2B dense transformer, following Qwen3-VL's image and video messaging conventions. The key to cross-embodiment support lies in action representation: a vehicle's ego pose, a camera's motion, and a robot arm's end-effector pose plus gripper state are unified into compact geometric vectors describing translation, rotation, and manipulation state, connecting control directly back to the visual structure of the scene.Source: MarkTechPost source · GIGAZINE source

Model S autonomous-driving death case ruled $240 million in damages, three state attorneys general seek to overturn · autonomy

Tesla lost a lawsuit over a pedestrian death involving Model S's autonomous driving system, with damages ruled at approximately $240 million; three state attorneys general subsequently sought to have the ruling overturned. Beyond the damages figure itself, what's worth watching is whether this becomes a reference point for allocating liability in assisted driving—as NHTSA is in the process of rewriting driverless vehicle rules around "capability" assessments, how courts allocate responsibility between human and machine will have a longer reach than any single payout.Source: IT Home (Chinese tech news outlet) source

Beijing's industrial robot output up 75.5%, service robot output up 2.3x in H1 · industrial

According to Beijing Municipal Bureau of Economy and Information Technology data: in H1, "AI+" initiatives drove output growth of 92.1%, 75.5%, and 2.3x respectively in integrated circuits, industrial robots, and service robots. Over the same period, the city's combined industrial and information software sector value-add approached RMB 1 trillion, or 36.6% of GDP, with fixed-asset investment in key industrial sectors up 24.0% year-over-year, including a 36.0% increase in high-tech manufacturing investment. The gap between service robots' more-than-doubled growth rate and industrial robots' rate suggests this wave of volume growth is landing more in commercial service applications than on production lines themselves.Source: Beijing News source

Kexing Shikong, Shanghai Technology Exchange, and Unitree launch real-robot data demonstration platform · adjacent

A jointly built high-quality real-robot data demonstration platform has launched, with Kexing Shikong's data collection team having already deployed some Unitree robots to Fuxing Island in Yangpu District, collecting data on real-world tasks such as chair repositioning in retail settings. The notable element here is bringing a technology exchange into the embodied data pipeline—once real-robot data needs to circulate across institutions, pricing and rights allocation become unavoidable, an area that has largely been a blank slate until now.Source: Shanghai Technology Exchange source (WeChat, CN) · Kanchai.com (Chinese tech news outlet) source

Booster Robotics launches flagship humanoid Booster T2, powered by Nvidia Thor · humanoid ⚠️ manufacturer's own claims

Booster T2 debuted at WAIC, powered by Nvidia's Thor chip, with the manufacturer claiming 2,070 TFLOPS of compute; the company also launched a development toolkit, Booster Studio. The company's stated goal is moving humanoids from "can move" to "can be used"; no mass-production or delivery figures have been disclosed yet.Source: Daofengti (Chinese tech media outlet) source (WeChat, CN)

XPeng unveils TuringViT vision encoder · autonomy ⚠️ manufacturer's own claims

XPeng has unveiled a vision encoder for physical AI applications called TuringViT, claiming it restructures the training paradigm for large vision models. Only the launch announcement is public so far, with no third-party reproduction or benchmark comparison yet seen.Source: Gasgoo (Chinese automotive industry outlet) source

Hardware · Supply Chain

· Lingyi iTech · full-machine contract manufacturing: Beijing Yizhuang's embodied intelligence super-factory targets stable annual capacity of 10,000 humanoid units in 2026, rising to 20,000 in 2027; the Chengdu facility has already stabilized mass production of AGIBOT-customized units, while the Beijing plant handles Tiangong-series mass-production validation ⚠️ investor-communication claims source (WeChat, CN)

· Data gloves · a collection-side bottleneck: mainstream Manus gloves cost RMB 60,000–100,000 per set and prices are still rising while supply remains scarce; pairing with an RMB 100,000-class dexterous hand pushes per-set collection cost toward RMB 200,000–300,000, prompting BrainCo (a 21-DOF exoskeleton, full set RMB 160,000), Dexrobot, and Linjiedian to each build their own gloves source (WeChat, CN)

· Visuo-tactile sensing · lifespan is a real problem: once the silicone skin tears or glue leaks out, the entire fingertip's data is ruined and the part must be replaced and recalibrated; Linjiedian says it has reached a lifespan of 100,000 grasps, while some manufacturers report needing to recalibrate after just two to three days of intensive data collection source (WeChat, CN)

· Weihong Corporation · micro servo motors: claims its dexterous hand's micro servo motors overcome the stalling and burnout issues of coreless motors, with projected monthly output of 80,000 units by year-end ⚠️ manufacturer's own claims source (WeChat, CN)

· China Aviation Optical-Electrical Technology: has joined the Ministry of Industry and Information Technology's humanoid robotics and embodied intelligence standardization working group source (WeChat, CN)

FutureX · Physical AI Daily — Issue 64 (07/21)

Shawn — Mon, 20 Jul 2026 14:46:52 +0000

Today's Highlights

· Blackstone invests about $676 million in South Korean actuator maker Futronic, betting on the robotics "components layer"

· China's Ministry of Industry and Information Technology: China has over 400 models of humanoid robots, about half the global total; quadrupeds account for nearly 70% of global sales

· Xiaomi releases Xiaomi-Robotics-1: a VLA model trained on 100,000 hours of real-robot data, pushing RoboCasa365 SOTA to 57.6%

· Yuequan Bionics (Chinese bionic robotics startup) raises a Pre-A+ round worth hundreds of millions of RMB; bionic dexterous hands reach 10,000-unit production scale with orders exceeding RMB 100 million

· UBTECH's humanoid robots to enter a South Korean auto-parts production line for a POC trial in August, marking embodied AI's overseas push shifting toward supply-chain integration

I. Research Progress

Xiaomi-Robotics-1: Scaling VLA with 100,000 Hours of Real-Robot Data · vla

Xiaomi has pushed VLA "data scaling" to a new magnitude — pretraining on 100,000 hours of real-robot manipulation trajectories (collected via UMI devices), paired with an automated annotation pipeline that labels each trajectory with natural-language descriptions of scene state transitions, providing fine-grained conditioning for action learning. The method splits into pretraining and post-training stages, showing a clear scaling curve: bigger models and more data yield better out-of-the-box real-robot performance, with the advantage transferring directly to post-training. It sets a new SOTA on RoboCasa365 with a 57.6% success rate (previous best: 46.6%), and averages 20.07 on RoboDojo (previous: 13.07); code and weights will be open-sourced.

Xiaomi Robotics Team · arXiv 2607.15330 source · Signal HF↑45

Harness VLA: Wrapping a Frozen VLA in a "Scheduling Shell" Instead of Retraining It · vla

A team led by Yu Chao at Tsinghua University proposes: rather than continuing to scale models and stack more robot trajectories, borrow from coding-agent design and give VLA its own harness. The frozen VLA is wrapped as a primitive handling only rich-contact actions like grasping and constrained placement, while contact-free movement, navigation, and release are handled by deterministic controllers; a planner then decides when to invoke the VLA, how to pre-position before invoking it, and how to retry after failures. With the same frozen VLA and unchanged parameters, simply changing how it is organized and invoked raised the success rate on the LIBERO-Pro generalization benchmark from 50.0% to 82.4% (compared with Nvidia's Cap-X at only 18.2% in the same period).

Yu Chao's team (Tsinghua University) · arXiv 2607.08448 source · Analysis: Embodied AI Insider source (WeChat, CN)

Orbis 2: Splitting Driving World Models into "Coarse Structure + Fine Frames" · world-model

Current world models mostly pursue frame-level fidelity at a single level of abstraction, lacking spatial and semantic reasoning. Orbis 2 splits future prediction into two levels: a high-level predictor forecasts coarse-grained scene structure over long horizons, and a low-level generator produces detailed frames based on that structure, balancing fidelity with reasoning. It also proposes a two-stage paradigm — "diffusion-forcing pretraining + teacher-forcing fine-tuning" — combining the representation quality of the former with the autoregressive stability of the latter, achieving SOTA across a suite of driving world-model benchmarks including long-horizon generation and counterfactual steering response.

Mittal et al. (University of Freiburg) · arXiv 2607.15898 source

EmArm: Giving Robotic Arms "Whole-Arm Human-Like Touch," Published in Nature Sensors · perception

By combining a rigid skeleton with flexible tactile skin, researchers built EmArm, a robotic arm with whole-arm human-like tactile perception that enables real-time physical adaptation — stably completing rich-contact manipulation while safely replanning trajectories in dynamic, unstructured environments. Compared with tactile sensing limited to the fingertips, it extends the "sensing zone" from the hand across the entire arm, pointing toward a contact-interaction approach closer to human touch.

Tang et al. · Nature Sensors source

Think at 5 Hz, Act at 20 Hz: Keeping LLM-Based Driving in Step with Control Frequency · autonomy

When large models are introduced into end-to-end driving, inference latency has always lagged behind the control frequency vehicles require, forcing existing approaches to invoke the model only every other cycle and reuse old commands in between — effectively ignoring the latest observation half the time. This paper proposes an asynchronous fast-slow architecture: a slow channel "thinks" at 5 Hz while a fast channel "acts" at 20 Hz, so every control step uses the latest frame. Open-loop waypoint error drops to nearly one-quarter of that from a backbone's built-in action head, with per-step model overhead of 32 ms independent of history length, runnable on a single consumer-grade GPU.

Li et al. · arXiv 2607.15621 source

Handroid: Combining "Dexterous Hand" and "Humanoid" into One Desktop Robot · manipulation

Dexterous hands and humanoids are typically developed separately — the former for object-scale rich-contact manipulation, the latter for locomotion and whole-body interaction. Handroid combines both into a single reconfigurable desktop-scale dual-morphology robot, uniformly supporting hand teleoperation, dexterous grasping, in-hand manipulation, and humanoid gait generation — providing a low-cost testbed for studying "hand × whole-body" coordination on one platform.

Li et al. · arXiv 2607.16187 source

Other papers today: AC-VLA uses compositional learning to strengthen out-of-distribution action execution, delivering roughly a 28% absolute improvement on compositional generalization tasks (arXiv 2607.15714 source); IMBench proposes an "intuitive manipulation" benchmark spanning perception—physical reasoning—action—iterative execution (arXiv 2607.15641 source); VTAP Gripper performs in-hand manipulation with a visuo-tactile active palm plus reconfigurable fingers (arXiv 2607.15448 source); MotionForesight adapts video models to predict future 3D scene flow from monocular video (arXiv 2607.16192 source); VTLoc uses 3D point clouds for tactile contact-point localization (arXiv 2607.16146 source); OptiTac uses fiber optics to mimic "artificial nerves" for bionic tactile skin (arXiv 2607.15746 source).

Open Source · Tools · Benchmarks

· Kuavo (X Humanoid): released what it calls the "world's first" large-scale cross-embodiment dataset and simulation engine, aimed at stable deployment of robots in real industrial scenarios source

· DexForce: open-sourced the Aligned DexWorld dataset, aligning dexterous manipulation data source

· Kunlun Wanwei (Chinese AI company) Matrix Game-3.5: open-sourced a game world model, focused on real-time responsiveness moving from "video generation" toward an "interactive world" source (WeChat, CN)

II. Funding & Deals

Futronic (South Korea) | Controlling-stake acquisition | About $676 million | Valuation of about KRW 1 trillion · hardware

Blackstone's private equity fund signed a definitive agreement to invest in Futronic, a South Korean maker of high-precision actuators and motion control systems, with founder Ko Jin-ho remaining as chairman and CEO — this is not a spin-off but a bet on expanded capacity. Founded in Busan in 1993, Futronic supplies actuators and motors for automotive and industrial robots, employs about 225 people, and is a company with three decades of experience in motion hardware. Rather than backing a front-facing humanoid brand, Blackstone is buying directly into the components layer that lets robots' "joints move and sustain load" — a so-called "sell the shovels" strategy: whichever full-robot maker wins, they will all need actuators.Source: Startup Fortune (citing Reuters/Bloomberg) source

Yuequan Bionics (Chinese bionic robotics startup) | Pre-A+ round | Hundreds of millions of RMB · hardware ⚠️ Order figures per company's own statement

Investors including Changfa, Huakong, Huaxia, Guoli Minsheng, and Xingke Venture Capital jointly invested, with existing shareholder Zhongguancun Qihang following on, marking the company's fourth funding round. Yuequan was founded by Ren Lüquan, an academician of the Chinese Academy of Sciences, and Ren Lei, a professor at the University of Manchester, pursuing a "bionic tension-compression structure" approach: self-developed magnetic-collector-drive artificial muscles integrate actuation, speed variation, and transmission into one unit, mimicking the human bone-muscle-tendon structure in an attempt to move beyond the conventional "motor + reducer + rigid joint" approach. Its flagship product, the Y-Hand M1 dexterous hand, has 38 degrees of freedom, full-hand tactile sensing, a grip strength of 28.7 kg, and single-finger repeat positioning accuracy of 0.04 mm. The company states its dexterous hands have reached 10,000-unit production scale with cumulative orders exceeding RMB 100 million.Source: 36Kr source (WeChat, CN) | Jingguan Ganzhi source (WeChat, CN)

Dobot (Chinese collaborative robotics maker) | Equity dispute alleged by co-founder ahead of ChiNext IPO hearing · industrial ⚠️ Dispute per the accusing co-founder's account

Collaborative robot maker Dobot faces an equity dispute alleged by a co-founder just ahead of its ChiNext IPO listing-committee hearing, with the company's flurry of recent funding rounds also drawing scrutiny. On the other hand, the company disclosed that first-half revenue and gross profit nearly doubled, with its collaborative-robot business becoming a "cash cow" underpinning its push into embodied AI commercialization.Source: Huxiu source

TÜV (Tuosida, Chinese industrial robot maker) files again for Hong Kong IPO, focused on integrated industrial robotics · industrial

Tuosida, whose core business is integrated industrial robots and self-developed R&D, has filed a second time for a Hong Kong listing, continuing the recent wave of Chinese industrial robotics companies rushing to list in Hong Kong during the current listing window.Source: 10jqka source

III. Commercialization & Deployment

UBTECH's Cruzr Y1 and Walker S2 to Enter a South Korean Auto-Parts Production Line · humanoid ⚠️ Still at POC validation stage

UBTECH signed a memorandum of cooperation with South Korea's Doksan JMR and AJIN Industrial to launch a production-line proof-of-concept (POC) for its Cruzr Y1 and Walker S2 humanoid robots at AJIN's newly opened manufacturing innovation factory in August 2026. Unlike its domestic-China deployments directly inside vehicle assembly plants for on-the-job training (Dongfeng Liuzhou Motor, Zeekr), this overseas move deliberately targets an auto-parts supplier, adopting a "UBTECH core technology + local Korean service" model to sidestep trade and compliance barriers. Whether the POC can convert into bulk orders remains to be seen, and local players like Rainbow Robotics are also accelerating their own efforts — this is one example of Chinese embodied AI's overseas expansion shifting from "exporting complete robots" to "localizing within the supply chain."Source: Netease Auto/Wangtongshe source

Jiushi Intelligence (Chinese autonomous-driving company): L4 Map-Free Autonomous Driving System Announces Mass Production · autonomy ⚠️ Per the company's own statement

At WAIC, Jiushi Intelligence stated its L4-level, HD-map-free autonomous driving system has entered mass production, emphasizing a low-cost, fast-deployment map-free approach.Source: Sina Finance source

YingShen's Embodied Robots Enter a Vietnamese Factory · industrial

YingShen's embodied robots have been deployed on a factory production line in Vietnam, marking another step in Chinese embodied AI's expansion into Southeast Asian manufacturing settings.Source: Vietnam Investment Review source

IV. Industry News

China's Ministry of Industry and Information Technology: China Has Over 400 Humanoid Robot Models, About Half the Global Total · humanoid

At a State Council Information Office press briefing on July 20, Wang Weiming, chief engineer of China's Ministry of Industry and Information Technology, shared first-half figures: China-developed quadruped robots account for nearly 70% of global sales share, and China has over 400 models of humanoid robots — more than half the global total. This follows Xi Jinping's appearance at WAIC setting the tone for "AI moving into the physical world" (previously reported), marking another official endorsement of the embodied AI industry's scale; WAIC also closed in Shanghai the same day. The same briefing also disclosed: China's new shipbuilding orders account for 81.2% of the global total, AI visual inspection has reached micron-level defect detection, and industrial large models have shortened preclinical R&D cycles for anti-tumor targeted drugs by 40%. The fact that over half of all model variants are Chinese-made indicates the "hardware supply" side is already highly saturated, with industry focus shifting from "whether it exists" to "who can deliver at scale and get robots actually working."Source: cnii.com.cn source (WeChat, CN) | Xinhua source

Unitree Releases UnifoLM-OminiA-0.3 Home-Care Embodied Model · embodied ⚠️ Per demo footage

Deployed on the G1, the model emphasizes a "single unified architecture," folding navigation, grasping, visual recognition, voice interaction, whole-home device coordination, and safety avoidance into one model — replacing the traditional layered architecture of separately calling perception/language/control APIs. In official demos, the G1 moves between living room, bedroom, and kitchen following a single spoken instruction, completing long-sequence tasks such as carrying a pillow, sorting dirty laundry, loading a dishwasher, and pulling a specific pill box out of a stack. The version number is deliberately marked 0.3, with technical details not yet disclosed; Unitree itself acknowledges the model is "somewhat slow to react, with occasionally clumsy grasping postures." This marks the fourth iteration in Unitree's UnifoLM series over the past year (world model → VLA → industrial → home care).Source: Deep Blue Embodied Intelligence source (WeChat, CN) | STCN source (WeChat, CN)

Fourier Intelligence (Chinese humanoid robotics company) Releases First Wheeled Dual-Arm Robot, GRW · industrial ⚠️ Per manufacturer specs

The brand's first wheeled-configuration robot targets flexible mobile assistance in factories — line-side delivery, tote handling, machine-side loading/unloading, palletizing and depalletizing, and other auxiliary stations that have long lacked automation. Specs: 29 degrees of freedom, dual-arm payload of 16 kg, peak torque of 500 N·m, repeat positioning accuracy of ±0.1 mm, a shoulder width of just 585mm to fit standard factory aisles, hot-swappable battery support for 24/7 continuous operation, and a no-code (natural language + drag-and-drop) tool chain that lowers the deployment threshold to accessory level. Fourier's reasoning: most factories have flat, hardened floors, so bipedal locomotion's terrain advantage goes unused while costing more in energy and stability — hence choosing a wheeled design to fill the automation gap at auxiliary stations first.Source: Robotics Grand Lecture Hall source

Hyundai Reportedly Moving to Further Tighten Control Over Boston Dynamics · humanoid

Following Hyundai's earlier $325 million acquisition of SoftBank's remaining 9.65% stake in Boston Dynamics (previously reported), South Korean media report that Hyundai Executive Chair Chung Euisun's personal stake in Boston Dynamics has risen to about 25%, as the group further consolidates control, paving the way for Boston Dynamics to be fully folded into Hyundai's robotics portfolio.Source: 매일경제 (Maeil Business Newspaper) source

Tesla to Report Q2 Earnings Wednesday, With Robotaxi and Optimus Progress in Focus · autonomy

The market is watching whether Tesla can demonstrate that its AI/Robotaxi/Optimus investments are "strengthening the company," with roughly 480,000 vehicles delivered in Q2. Separately, according to NHTSA data, Electrek reported that a Tesla remote operator was linked to a Robotaxi collision in Houston, adding further controversy over the safety of its remote-teleoperation setup.Source: Investor's Business Daily source | Electrek source

Hardware & Supply Chain

· Wolong Electric Drive × Shangwei New Materials: reached a partnership at WAIC combining motor drives with composite materials, entering upstream joints and structural components for humanoid robots source

· Solid-state batteries: industry discussion is increasingly framing their economics as a key variable for humanoid robots' battery life and safety source

FutureX · Physical AI Daily — Issue 63 (07/20)

Shawn — Sun, 19 Jul 2026 14:44:01 +0000

Today's Highlights

· ModelBest (Chinese AI lab spun out of Tsinghua) open-sources embodied model MiniCPM-Robot, crushing π0.5's memory-benchmark score of 10 with 53 points

· The founder of Huawei's embodied-brain division departs to start his own company; Cognibrain, positioned against JEPA, raises a nine-figure RMB round

· China's "Measures on Anthropomorphic AI Interaction Services" take effect, pulling emotional-companion robots into tight regulation

· Former head of DJI's robot vacuum business pivots to exoskeletons; Sequoia China and Shunwei Capital back Yueshan Innovation (OROX)'s angel round

· WAIC organizers disclose: China's core AI industry has surpassed RMB 1.2 trillion in scale

1. Research Progress

Universal transferable visual attack: a single adversarial patch defeats VLA models across the board · vla

As VLA (vision-language-action) models become the dominant "brain" for robots, the robustness of their visual input pipeline is emerging as a security weak point. This work by a Nanyang Technological University team, named a CVPR 2026 best paper, produced an adversarial patch that is universal and transferable across models — placed within a camera's field of view, it induces different VLA models' action policies to fail, with no need to tune it for each model individually. It's a reminder that when robots actually do physical work, "seeing one thing wrong" can lead to consequences far more severe than a misclassified image.

Nanyang Technological University team · CVPR 2026 best paper · Analysis: Ru-hulu的具身智能分享会 (Ru-hulu's Embodied Intelligence Sharing) source (WeChat, CN)

Open Source · Tools · Benchmarks

· MiniCPM-Robot (ModelBest × OpenBMB): ModelBest, a Tsinghua-linked AI lab, made its WAIC debut in embodied intelligence, open-sourcing two on-device models. The general-purpose VLA MiniCPM-RobotManip, trained on the multimodal on-device model MiniCPM-V 4.6, has just 1.5B parameters yet matches 3–4B-parameter models — by aggressively compressing visual tokens it packs historical observations into inference, achieving 1-minute native memory. On RMBench, which tests contextual memory, it scores 53 versus π0.5's 10 (near-random). At 60 frames of history, per-step inference drops from a traditional 125 TFLOPs to 3.3 TFLOPs, with single-step latency of 120ms on an H100 — about half that of π0.5. A second, 0.9B model, MiniCPM-RobotTrack (co-developed with Nanjing University), is built for offline local tracking, natively supporting Unitree's Go2 — it can track a person at 5+Hz even with the network cable unplugged, and posts scores of 89.8/73.4/80.4 across three tracking tasks, an open-source SOTA. Both are available on GitHub/Hugging Face. source (WeChat, CN)

· PhyAI (ModelBest): An embodied reasoning framework open-sourced alongside the models, built for fast on-device inference and large-scale cloud RL rollouts; the company claims speedups of roughly 2.85×/1.82×/2.28× for π0/π0.5/GR00T respectively on an RTX 5090. source (WeChat, CN)

2. Funding & Deals

Cognibrain (具脑磐石) | New round | Nine-figure RMB · world-model

The round was led by industrial capital with backgrounds in brain-inspired computing and embodied intelligence, with follow-on investment from existing shareholders; Duowei Capital served as exclusive financial advisor, and another financing round is closing in parallel. Founded in 2025, the company's founder and CEO Zhu Senhua was formerly "the top person for Huawei's embodied brain division," having served as director of Huawei Cloud's AI Algorithm Innovation Lab and led the Pangu embodied foundation model; he holds a PhD in cognitive neuroscience from the University of Pennsylvania. The team is pursuing a brain-inspired intelligence approach to build a "cognitive world model," aligned in direction with Yann LeCun's JEPA but emphasizing a full closed loop from "representation-prediction" through to "decision-execution," with a focus on low data requirements, high generalization, lifelong learning, and low power consumption. This marks another team with big-tech embodied-AI roots striking out independently, as the industry's buzzword shifts from "VLA" to "world models."Source: 36氪硬氪 (36Kr Hard Tech) source

Yueshan Innovation (OROX) | Angel round | Amount undisclosed · adjacent

Co-invested by Sequoia China and Shunwei Capital. Founder Chen Liangshen previously led DJI's supply chain and robot vacuum business, where he headed the launch of DJI's first sweep-and-mop robot, ROMO, which sold over 10,000 units in its first month. The company builds consumer-grade exoskeletons, applying DJI's "R&D–supply chain–distribution" playbook to a category still stuck between lab prototypes and mass production. A notable reference point: online retail sales of intelligent walking-assist exoskeletons in China rose 785% year-over-year from January to April this year, with retail prices falling from hundreds of thousands of yuan to the thousand-yuan range. This also reflects a shift in early-stage investing — teams with algorithm patents but no manufacturing and delivery capability are no longer the top pick.Source: 具身涌现 (Embodied Emergence) source (WeChat, CN)

Coram AI | Series B | $35 million · adjacent

Led by Ansa Capital and Battery Ventures. Coram builds AI video security software, connecting existing cameras to models for real-time understanding and alerts — a perception/monitoring-layer player adjacent to robotics.Source: ascendants.in source

Wildfire Robotics (山火机器人) | Angel round | Tens of millions of RMB · hardware

Led by Haiyuan Capital, with proceeds going toward team expansion and R&D. The company focuses on localizing production of core transmission components for humanoid robots in China, targeting the reducer/transmission segment — currently the tightest bottleneck upstream.Source: 链榜 (Chain Ranking) source (WeChat, CN)

3. Commercialization & Deployment

Honda launches commercial autonomous electric lawn mower, priced at $32,999 · autonomy

Honda has launched an autonomous electric lawn mower for the commercial market, priced at $32,999. Compared with passenger-carrying autonomous driving, lawn maintenance is a closed environment with low failure costs — putting L4 capability first into "no passengers, fault-tolerant" work machinery is a low-profile but genuinely scaling branch of autonomous-driving commercialization.Source: 新浪网 (Sina) source

Aomi Technology (OYMotion)'s dexterous-hand orders top 10,000 units in H1 · embodied

Dexterous-hand maker Aomi Technology (OYMotion) disclosed that orders surpassed 10,000 units in the first half of the year. In a WAIC week saturated with dexterous-hand "demos," this is one of the few figures that reflects real order volume; set against this year's price collapse for dexterous hands — from "100,000+ RMB" down to the "10,000 RMB" range — a rising volume number is worth more attention than spec sheets.Source: 新浪财经 (Sina Finance) source

San Francisco blackout briefly halts Waymo robotaxi service · autonomy

A major power outage in San Francisco briefly disrupted Waymo's robotaxi service and paused freeway routes, before service resumed following adjustments. Scaled autonomous vehicle fleets depend heavily on charging, dispatch, and communications infrastructure — a city-wide blackout serves as a real-world stress test of their redundancy and recovery capability.Source: 디지털투데이 (Digital Today) source

4. Industry Developments

Regulators convene closed-door meeting; emotional-companion robots pulled into tight regulation · humanoid

According to 21st Century Business Herald, regulators convened a closed-door meeting with leading companies last week on emotional-companion robots, focused on data security, cybersecurity, and standards for emotional interaction. The "Interim Measures on Anthropomorphic AI Interaction Services," effective July 15, has now formally taken effect, bringing all AI products/robots that provide sustained emotional interaction through anthropomorphic forms under tight regulation, requiring hard compliance measures including protections for minors, addiction-prevention features, content moderation, and extreme-emotion intervention. The category is red-hot: UBTech's U1 hyper-realistic humanoid sold more than 10,000 units in a single day, starting at RMB 119,800 and topping out at RMB 990,000, for an estimated transaction value of nearly RMB 2.2 billion (per brokerage estimates). But sustained companionship requires long-term collection of voice, imagery, and even private emotional records, magnifying data-leak and ethical risks — this "trillion-yuan opportunity" remains a chaotic category with unresolved gaps in data security, compliance, and user experience.Source: 21财经 (21st Century Business Herald) source

Official WAIC data: China's core AI industry surpasses RMB 1.2 trillion · adjacent

The 2026 World Artificial Intelligence Conference disclosed that China's core AI industry has surpassed RMB 1.2 trillion in scale, with "embodied intelligence going to work in factories" named as this year's most distinctive industry signal; organizers also stated that roughly 120,000 high-quality datasets have been built. Beyond the numbers, the notable shift is in narrative — a consensus emerged at this year's WAIC that the focus is moving from showcase demos to real deployment in factories and homes.Source: 新浪财经 (Sina Finance) source

Da Xiao Robotics unveils "Kaiwu World Model" Kairos 3.1 · world-model ⚠️ Vendor claim

Da Xiao Robotics (ACE Robotics) launched Kairos 3.1 (Kaiwu World Model 3.1) at WAIC, touting "native action integration" and claiming to be the world's first embodied world model driving a robot body via on-device execution. Both the "world's first" claim and the on-device real-time performance are the vendor's own statements; real-world generalization and deployment performance await third-party verification.Source: 新浪财经 (Sina Finance) source

Shouqi Ride Hailing launches SAIC robotaxi mass-production customization project, debuting in 2027 · autonomy

Shouqi Ride Hailing, together with SAIC Motor and a leading autonomous-driving technology provider, has launched a robotaxi mass-production customization project. Officials state it will be Shanghai's first factory-built, mass-produced, purpose-designed vehicle for high-level autonomous driving, with the new vehicle slated to debut in 2027. This remains at the planning stage and deployment is still far off, but an automaker directly defining a factory-built, mass-produced robotaxi is another example of OEMs bringing robotaxis into their core business.Source: 上观新闻 (Shanghai Observer) source

Google DeepMind: video generation models already "contain" the world models computer vision has been missing · world-model

Amid ongoing industry debate over the boundaries of "world models," Google DeepMind published an argument that today's video generation models already implicitly contain world models — the piece long missing from computer vision, they argue, can be "read out" of video generators. This is a research stance rather than a settled conclusion, but it resonates with this week's broader push from academia and industry to redefine "what counts as a world model," pulling the debate back from terminology toward the question of what these models have actually learned.Source: the-decoder.com source

Hardware · Supply Chain

· Weihong Co. / Changzhou Hanmu: Holds the top market share in dexterous-hand motors, with monthly shipments of around 18,000 units; reportedly a co-development partner for many leading dexterous-hand companies, with supply set to ramp up in the second half of the year. source (WeChat, CN)

· 9.9mm frameless torque motor: A company has built the world's smallest 9.9mm frameless torque motor, along with a 3.5g fingertip motor for dexterous hands — pointing toward hands with higher degrees of freedom and dimensions closer to a human hand's "fingertip heart." source (WeChat, CN)

· Dexterous-hand price drops: Robotic dexterous hands once priced at "100,000+ RMB" have now fallen into the 10,000 RMB range, and this price collapse is drawing in new entrants such as cross-border e-commerce players. source (WeChat, CN)

FutureX · Physical AI Daily — Issue 62 (07/19)

Shawn — Sat, 18 Jul 2026 14:42:50 +0000

Today's Highlights

· Tactile-sensing company Yimu Technology completes Series E round exceeding RMB 1 billion, valuation tops RMB 10 billion to become new unicorn

· Shanghai's state-industry co-built humanoid pilot production platform launches, with an annual capacity of 2,000 full units and 40+ signed clients

· Zhiyuan Robotics declares "Year One of Deployment" at WAIC, with 60 robots on duty providing public services at the venue

· RoboScience's cloud-based embodied foundation model demonstrates a live 30-second hand swap, one brain controlling 10+ dexterous hand models

· The "world model" concept proliferates; at WAIC, academia and vendors collectively try to "set the record straight" on its meaning

I. Research Progress

ScaleBFM: Disentangling the Scaling Laws of Humanoid "Behavior Foundation Models" · locomotion

Humanoid whole-body control has long lacked a clear "bigger is better" curve like the ones seen in language models. ScaleBFM scales training data to 102 million frames and, for the first time, independently measures three separate dimensions — motion data volume, PPO sample volume, and model capacity — offering an actionable scaling recipe for where to direct resources in humanoid "Behavior Foundation Models" (BFMs). The work comes from a team including researchers at the Chinese University of Hong Kong.

Chinese University of Hong Kong et al. · arXiv 2607.15163 source · Analysis: 具身智能研究室 source (WeChat, CN)

FreMotion: Using "Frequency Decoupling" to Break the Precision-Smoothness Dilemma, a 100M-Parameter Model Outperforms a 3.3B VLA · manipulation

The experiments reveal a counterintuitive fact: even a 3.3-billion-parameter, π0-class VLA sees its success rate collapse in fine-grained manipulation tasks lasting longer than 10 seconds, because vision and action are treated as a homogeneous information stream. FreMotion decouples high-frequency visual detail from low-frequency action intent, and with only 100 million parameters outperforms the larger model on fine manipulation. The paper appears in IEEE T-ASE.

Analysis: 具身智能制造 source (WeChat, CN)

CosFly-VLA: Adding 3D Spatial Priors for Drones, Recovering Targets Even After Occlusion · autonomy

General-purpose VLAs like OpenVLA "follow what they can see" but lack drone-specific altitude and distance geometric priors, so errors accumulate under prolonged occlusion and the target is eventually lost. CosFly-VLA introduces spatially-aware representations, enabling "occluded pedestrian tracking" in urban environments to precisely recover the target after it disappears — pointing to a real-world shortfall of VLAs in low-altitude scenarios.

Analysis: 视觉语言导航 source (WeChat, CN)

Open Source · Tools · Benchmarks

· Damo Academy Digital Teleoperation Data Collection Scheme: Alibaba's Damo Academy has open-sourced a "digital teleoperation" embodied data collection scheme that lowers the barrier for real-robot demonstration, targeting data scarcity — a core bottleneck for embodied AI deployment source (WeChat, CN)

· LeRobot Extension: NVIDIA and Hugging Face jointly add new tooling to the open-source robotics framework LeRobot, continuing to push the open-source embodied AI toolchain forward source

II. Funding and Deals

Yimu Technology (Chinese tactile-sensing startup) | Series E | Over RMB 1 Billion | Valuation Tops RMB 10 Billion · hardware

Yimu Technology, focused on robotic tactile perception, announced completion of a Series E round exceeding RMB 1 billion, with valuation surpassing RMB 10 billion, making it a new unicorn to emerge during WAIC; the round was jointly participated in by several leading RMB funds, top-tier USD funds, and industrial partners, with over ten cumulative funding rounds to date — Shunwei Capital and TCL Capital have both invested across multiple rounds. The company's flagship product is a biomimetic visual-tactile sensor, now iterated down to under 3 millimeters thick with force resolution of 0.005N, and claims 90% of its core tactile metrics reach human-like levels. What propelled the valuation from single-digit billions to over RMB 10 billion is not revenue but the market's repricing of its role as a "physical AI data gateway" — the company is simultaneously pushing VTLA (Vision-Tactile-Language-Action) and the open-source tactile dataset TouchNet, aiming to control this tactile data spigot. Founder Li Zhiqiang holds a PhD from Carnegie Mellon.Source: Robo百科 source (WeChat, CN)

Beijing Natural Will Technology | Angel Round | Amount Undisclosed · embodied

Natural Will Technology, founded by a Tsinghua PhD advisor, completed an angel round with Ant Group and CATL joining Sequoia China, Cathay Capital, Fountainvest, and Huakong Dianke; disclosures indicate the company secured three funding rounds within six months of registration, continuing the industrial-capital playbook of "binding to a core team, betting heavily and early."Source: 具身涌现 source (WeChat, CN)

Rushen Robotics | Pre-A Round | RMB Hundreds of Millions · embodied

Rushen Robotics, focused on embodied elder-care services, completed a Pre-A round in the RMB hundreds of millions (closed in late April), as an alumnus company of Miracle Plus's 2023 spring cohort; elder care is an embodied AI use case with clear demand but extremely high requirements for safety and compliant manipulation.Source: Miracle Plus source (WeChat, CN)

Zhuizhi Gongke | Seed Round | RMB Tens of Millions · industrial

Zhuizhi Gongke completed a seed round of RMB tens of millions, applying industrial agents to reshape materials processing so robots can learn "feel" in processes such as polishing; industrial robotics funding has been relatively cold over the past two years, so a seed round targeting a specific niche process signals industrial-side bets on "embodied AI entering the factory."Source: RecodeX source (WeChat, CN)

UBTECH | Hong Kong Share Placement | HKD 436 Million · humanoid

UBTECH plans another share placement of approximately HKD 436 million, its second such placement in under a year since its Hong Kong listing, to fund ammunition for mass production and delivery of humanoid robots — the company had previously disclosed that orders for its Walker S and U-series robots already exceed ten thousand units, straining cash consumption in tandem with capacity ramp-up.Source: Futu Niuniu source

III. Commercialization and Deployment

Zhiyuan Robotics' "Year One of Deployment": Five New Products + 60 Robots On Duty for Venue Services · humanoid

Under the theme "Year One of Deployment," Zhiyuan Robotics (Chinese humanoid robot maker) released five new products built around its "Three Intelligences in One" architecture, and deployed roughly 60 robots at the WAIC venue to handle public services such as guided tours and duty stations; its Yuanzheng A3 Ultra was named one of the conference's "Ten Treasures of the Venue," the only humanoid robot selected, and Thailand's prime minister also visited the booth. Zhiyuan is pushing toward being "deployable, replicable, and scalable," though 60 units still represents exhibition-scenario operation, and the pace toward large-scale commercial deployment awaits confirmation from actual orders.Source: 智元朋友圈 source (WeChat, CN)

JAKA (Dobot) × Tencent: Physical AI's "Shenzhen Solution" Runs End-to-End on a Real Production Line · industrial

Dobot Robotics and Tencent deployed a full-stack embodied AI solution on a real production line, demonstrating end-to-end operation from perception, decision-making to execution, packaged as a replicable Physical AI "Shenzhen Solution." Compared with booth demos, running the full chain at production-line scale is closer to actually "working in the factory," and represents one data point in this year's WAIC narrative shift "from performative demos to actual work."Source: Sina Mobile source

WeRide: Driverless Robotaxis Cover Three Emirates in the UAE, Selected as National-Level Overseas Expansion Case · autonomy

WeRide's fully driverless robotaxi service has expanded to cover three emirates in the UAE, and was named the only autonomous-driving company among WAIC's ten flagship cases, listed as a national-level model for AI's global expansion. The Middle East remains the largest-scale destination for Chinese robotaxi companies' overseas expansion, and the broadening of real overseas operating range carries more weight than opening a single new city.Source: Stock Titan source

Mech-Mind: 27,000 Embodied AI Deployments, Emphasizing "One Brain, Many Forms" Generality · industrial

Mech-Mind Robotics disclosed at WAIC roughly 27,000 cumulative embodied AI deployments, emphasizing a single general-purpose intelligence system that adapts to multiple robot bodies, including robotic arms and composite robots, so robots are no longer tied to a single task or object. Shipment volume in industrial manipulation is currently one of the few deployment metrics that can be backed by hard numbers.Source: 甲子光年 source (WeChat, CN)

JD.com × Zhiyuan Robotics' Genie G2 Max: Package Pickup, 24-Hour Hauling (Previously Reported) · industrial

The deployment of the JD Logistics × Zhiyuan Robotics Genie G2 Max at JD's "smart wolf warehouse" was previously reported; today's coverage adds demonstration details such as "first-ever package pickup" and "24-hour continuous hauling," with no new milestones.Source: 驱动之家 source

IV. Industry Developments

RoboScience's Cloud-Based Embodied Foundation Model Visics: Live 30-Second Hand Swap, One Brain Controlling 10+ Dexterous Hands · world-model

Making its first WAIC appearance, RoboScience (Chinese embodied AI startup) used a single "brain" — Visics — to drive over 10 dexterous hand models from different brands on stage, completing detachment, swapping, and regrasping in just 30 seconds, with no need for recollecting data or on-site training, which the company calls the industry's first "one brain, multiple hands, plug-and-play" system. Its VLOA architecture treats "how an object should be manipulated" (Object Trajectory, the continuous 3D point-cloud trajectory of an object) as a new token: an upper-level world model previews the object's trajectory, while a lower-level manipulation model translates it into actions for the current robot body, enabling reuse across different robot bodies, objects, and tasks. Visics is also delivered as pure software deployed to the cloud, which the company claims is the world's first cloud-deployable embodied foundation model (EaaS). It reportedly maintains a multi-SKU grasping success rate stable above 99% and has piloted with several retail and logistics companies. Both the hand-swap challenge and success rate figures come from live demonstration and vendor disclosure; scaled stability still awaits third-party verification.⚠️ Demonstration figuresSource: 机器之心 source (WeChat, CN)

Shanghai's State-Industry Co-Built Humanoid Robot Pilot Production Platform Launches, Annual Capacity of 2,000 Full Units · industrial

Led by the Shanghai Municipal Commission of Economy and Informatization, and jointly built by the National-Local Joint Humanoid Robot Innovation Center and Shanghai Electric, the Shanghai Humanoid Robot Pilot Production Service Platform has officially launched. Located in Pudong's Robot Valley, with a floor area of roughly 6,800 square meters, it houses two flexible full-unit pilot production lines with an annual capacity of 2,000 full units, plus capacity for 2,000 sets of joint/dexterous-hand core modules per year. The testing area includes equipment such as the industry's first comprehensive dexterous-hand test bench, covering 400+ verification items. The platform lowers the tens-of-millions-RMB barrier to pilot production for small and mid-sized teams through a shared public model, establishing unified assembly and inspection standards, directly addressing the gap between "hard-to-productionize prototypes" and "unstable mass production." Over ten companies have already signed on, including Zhuoyide (pilot production of thousands of units of its fully biomimetic robot Moya) and Zhenghe Industrial. This is government-and-industry-led infrastructure, not a single vendor's figures, and points to a common bottleneck in the "last mile" of mass production.Source: 人形机器人世界 source (WeChat, CN)

The "World Model" Concept Proliferates; Academia and Vendors Collectively "Set the Record Straight" at WAIC · world-model

"World model" became this WAIC's most frequently used — and loosest — label: teams doing video generation, 3D reconstruction, and robot simulation are all attaching the term to their own models, with multiple parties pointing out it is becoming a new catch-all buzzword following "metaverse" and "large models." Yao Maoqing, a partner at Zhiyuan Robotics, argued at the main forum that the essence of a world model is "accurately predicting the next state of the physical world," requiring multimodal fusion and causal understanding, framing it as moving "from virtual generation to physical truth-seeking"; Shen Yujun, chief scientist at Ant Group's Lingbo, and Huang Tiejun of the Beijing Academy of Artificial Intelligence, among others, drew the line by distinguishing world models from VLA's "statistical fitting," arguing they must genuinely understand physical laws. For industry readers, the key distinction is: being able to generate coherent visuals ≠ being able to reason about physical consequences — the latter is what embodied AI deployment actually needs.Source: 老冯云数 source (WeChat, CN)

Tencent Rounds Out Its Full Embodied AI Stack at WAIC: Three Hunyuan Foundation Models + Tairos Agents + ADP 4.0 Goes Overseas · embodied

Tencent released multiple embodied foundation models in one go (the Hy-Embodied series' "left brain, right brain") along with agents Apexio and TairosAgent, integrating perception, decision-making, and action into a continuous closed loop, and upgraded its full-stack embodied AI solution to ADP 4.0, launching an overseas version and a full upgrade of its "Tailuo" platform; the company said related AI call volume surged 68-fold year over year. Previously keeping a relatively low profile, Tencent used WAIC to round out its full-line capabilities across robotics, office productivity, and overseas expansion in one move, formally joining the competition for embodied foundation models.⚠️ Vendor figuresSource: Sina Finance source

Galbot (Chinese embodied AI startup) Releases WAM-TTT: Compressing Post-Training Cost from Teleoperation Down to "Phone-Recorded Video" · world-model

Galbot released at WAIC WAM-TTT, a test-time training framework for its World-Action Model, positioned as a key module for scaling deployment of its "Xingnao" system. The company claims it addresses the pain point of "generalization decaying whenever the environment changes upon deployment," compressing the data needed for post-training from real-robot teleoperation-level down to ordinary video-recording level. The WAM approach (fusing world models with VLA) was previously debuted jointly with Peking University at ICCV and is seen as one direction for next-generation embodied AI; the actual generalization gains remain vendor-demonstrated disclosures.Source: NetEase Mobile source

Joyson Electronics Releases Production-Grade "Brain-Hand-Capability" Full Stack, Debuts Dexterous Hand · hardware

Joyson Electronics, a leading auto-parts manufacturer, debuted at WAIC a production-grade "brain-hand-capability" full-stack solution along with a proprietary dexterous hand and semi-solid-state battery, stating that its robot controllers are already in mass-production supply; this kind of Tier 1 supplier bringing automotive-grade manufacturing and mass-production experience into embodied hardware is one of the main threads on the supply-chain side at this year's event.Source: Sina Finance source

Goldman Sachs: Humanoid Robots Becoming the "Next Monetization Frontier," Issues Buy Ratings on 5 Names · adjacent

A Goldman Sachs research report judges that the AI investment narrative is shifting from infrastructure to the application layer, stating directly that humanoid robots are becoming "the next clearest monetization frontier," and issuing buy ratings on 5 names; however, it also projects that widespread commercial deployment will not arrive until 2027–2029.⚠️ Institutional viewSource: 云知观点 source (WeChat, CN)

Hardware · Supply Chain

· AGLINK: Cumulative deliveries of its dexterous hand series exceed 8,000 units, ranking first in the global dexterous-hand market in Q1 2026; debuted its flagship Ultra-M at WAIC source (WeChat, CN)

· STMicroelectronics: Promoting highly integrated chips to lower humanoid robot costs, while targeting the South Korean supply chain source

· AgiBot Dexterous Hand × Shanghai Electric: Released industrial-grade five-finger dexterous hand DexHand021 Pro at WAIC, and established a joint venture with registered capital exceeding RMB 100 million (Series A previously reported) source

FutureX · Physical AI Daily — Issue 61 (07/18)

Shawn — Fri, 17 Jul 2026 15:07:47 +0000

Today's Highlights

· Xi Jinping attends WAIC for the first time and sets the tone for AI moving into the physical world, giving embodied intelligence its highest-level policy endorsement to date

· AgiBot's Genie G2 Max partners with JD Logistics — the first time a humanoid robot has entered a real production warehouse setting

· Zoox issues a software recall for its robotaxi fleet after a thick-smoke misjudgment, covering 105 vehicles

· Xynova (Chinese dexterous-hand startup) raises a RMB 500M Series A+ led by Meituan, pulling in RMB 1.5B in 7 months to become a new unicorn in dexterous hands

· OFweek: mainland China's humanoid robot production topped 30,000 units in H1, with shipments of about 23,000 units

I. Research Progress

RoboTTT: Stretching Robot Policy "Context" by Three Orders of Magnitude · vla

Robotic foundation models have previously only been able to consume single-step or short-history visual-action context. RoboTTT is the first to extend this window to 8K timesteps — three orders of magnitude longer than existing policies — without slowing down inference. The approach embeds test-time training (TTT) inside a VLA: a set of "fast weights" continuously undergo gradient updates during both training and inference, compressing long history into weight space for retrieval as needed. Overall success rate on real-robot manipulation improved 87% over a single-step baseline, and the model independently completed a five-minute, ten-stage assembly task that all baselines failed at; the 8K context outperformed a 1K pretraining context by a further 62%. The authors propose "context length" as a new scaling axis for robotic foundation models.

Yunfan Jiang et al. (NVIDIA GEAR) · arXiv 2607.15275 source · HF↑11

A Scaling Recipe for Humanoid "Behavior Foundation Models" · locomotion

Whole-body humanoid control demands natural coordination, real-time responsiveness, and cross-environment generalization all at once. The team revisited the scaling laws for behavior foundation models (BFMs) and found three factors must be co-optimized: unifying diverse controls into motion tracking under a global coordinate frame, the ratio between online rollout count and reference motion diversity, and a scalable Humanoid Transformer architecture. In simulation and on real hardware, keypoint pose error (MPKPE) dropped 82% under the global mode compared with existing controllers.

Weishuai Zeng et al. · arXiv 2607.15163 source

DriftWorld: Pushing Robot World Models to 30+ FPS with "Drift" · world-model

Using world models for planning requires generating large volumes of rollouts quickly, but diffusion-based world models are expensive due to multi-step sampling at every call. DriftWorld instead learns an "action-conditioned drift" during training, so at inference a single forward pass generates future frames — averaging 17x faster than diffusion baselines and reaching 30+ FPS, while achieving SOTA decision-making performance with less inference time on benchmarks like Bridge-V2, RT-1, and Push-T. When used as an offline "policy scorer," its scores correlate with real success rates at up to 0.99.

Susie Lu et al. · arXiv 2607.15065 source · Analysis: 世模人 source (WeChat, CN)

When World-Action Models "Think Right but Act Wrong": A New Class of Adversarial Attack · world-model

World-action models (WAMs) couple action generation with future prediction, and the field has generally assumed that "verifying actions against imagined futures" makes them safer. BadWAM disproves this assumption's robustness: tiny visual perturbations can decouple the alignment between "imagination" and "execution" — an aggressive mode drives closed-loop success rate straight down from 96.5% to 43.1%, while a stealth mode leaves the model still "imagining" plausible-looking futures while executing misaligned actions, bypassing imagination-based safety checks entirely.

Qi Li et al. · arXiv 2607.15207 source · HF↑34

RxBrain: Merging Language Reasoning and "Visual Imagination" into One Embodied Cognitive Base Model · world-model

Unlike VLMs focused on scene understanding, or generative world models that only predict future frames, Tencent Hunyuan's RxBrain represents embodied plans as a single sequence: language provides an abstract skeleton of task decomposition, constraints, and decision logic, while visual imagination uses world-state prediction to ground each step in concrete physical states. It employs a unified Mixture-of-Transformers, enabling understanding and generation across language/image/video within a single model; when extended to continuous action, it achieves usable real-robot performance without large-scale action-data pretraining.

Haotian Liang et al. (Tencent Hunyuan) · arXiv 2607.14187 source

Reflex: Streaming Inference Lets Flow-Matching VLAs Output While Reasoning, Achieving Real-Time Control · vla

Flow-matching VLAs offer high precision for continuous control, but iterative denoising is inherently at odds with real-time robotics: injecting global timesteps breaks KV-cache, forcing a choice between slow recomputation and "mathematically incorrect cache reuse." Reflex restructures this process with streaming inference, compressing latency to real-time-executable levels while preserving continuous-control precision.

Yuanchun Guo, Bingyan Liu · arXiv 2607.14695 source

MIDAS Hand: A $3,000, 3D-Printed Open-Source Dexterous Hand · manipulation

The bottleneck in dexterous manipulation isn't only algorithmic — it's also the lack of hardware that balances human-like hand size, ease of manufacture and maintenance, tactile sensing, and controlled cost. MIDAS Hand delivers a 16-DoF (13 active) open-source hand with direct-drive low back-drive torque and 283 integrated triaxial tactile units, weighing 700 grams total, with a bill-of-materials cost under $3,000 and 3D-printed parts assemblable within three hours — released alongside design files, control/tactile APIs, simulation models, and a teleoperation pipeline.

Alvin Zhu et al. · arXiv 2607.14487 source

Other papers today: Open-AoE, an open first-person manipulation dataset and toolchain (arXiv 2607.14183 source); Wan-Streamer v0.3, a real-time full-duplex audio-video interaction model framing "video = world + event stream" (arXiv 2607.15038 source); Instant NuRec, feed-forward 3D Gaussian reconstruction accelerating autonomous-driving scene simulation, from NVIDIA (arXiv 2607.14203 source); Never Too Late for Force, using reactive force injection to accelerate VLA post-training (arXiv 2607.14236 source); KineFuse, kinematics-aware tactile fusion for in-hand pose tracking of occluded objects (arXiv 2607.14842 source); Active Real-World Factor-Based Evaluation, an active real-robot factorized evaluation for general robot policies (arXiv 2607.14439 source); Beyond Visual Grasping, a complex grasping benchmark spanning detection to execution (arXiv 2607.14341 source).

Open Source · Tools · Benchmarks

· HY-World 2.0: Tencent Hunyuan's open-sourced systematic multimodal world model, running a four-stage pipeline through understanding—synthesis—reconstruction, claimed to match closed-source commercial models source (WeChat, CN)

· MoWorld 3D: the first collaborative world model released by China's National AI Application Pilot Base (Embodied Intelligence), developed jointly with Huawei Cloud compute source (WeChat, CN)

· EmbodiedCLUE-World 2.0: an updated release of an embodied world-model evaluation benchmark, led by Peking University's PF-Cosmos, emphasizing physical laws and executable manipulation workflows source (WeChat, CN)

II. Financing & Deals

Xynova (Chinese dexterous-hand startup) | Series A+ | RMB 500M | Reaches Unicorn Status · embodied

Led by Meituan, with follow-on from NIO Capital, China Merchants Capital, and an unnamed major internet company, plus additional backing from existing strategic investor Xiaomi. Xynova makes dexterous hands for humanoid robots; founded in late 2024, it has completed 4 funding rounds totaling roughly RMB 1.5B in under two years. Its flagship product, Flex 2, uses a hybrid "tendon-driven + direct-drive" architecture with 23 degrees of freedom, a total hand weight under 400 grams, and fingertip force exceeding 20N. The company has built a 5,400-sqm factory in Hangzhou and plans to reach annual capacity of 10,000 dexterous hands and 200,000 micro linear actuators by the end of 2026 — the shift from "building one hand" to "reliably replicating ten thousand hands" is exactly the threshold for competing to become a Tier 1 supplier. Industrial capital including CATL, Xiaomi, JD, Li Auto, Meituan, and NIO is placing dense bets on this "last centimeter." Source: NE时代智能体 source (WeChat, CN)

MoSense (Chinese tactile-sensing startup) | Angel Round | Tens of Millions RMB · embodied

Jointly invested by Sequoia China, Hillhouse Ventures, and AgiBot. MoSense builds whole-body multimodal fusion tactile systems for robots; founded in May 2026, headquartered in Shanghai with an R&D center in Shenzhen's Qianhai. Tactile sensing is the "skin" that connects intelligent decision-making to the physical world — top-tier VCs joining forces with industry players continues this week's capital heat in the tactile/dexterous-hand sector. Source: 机器人前瞻 source (WeChat, CN)

Kunlunxing Robot | Three Rounds in 90 Days Since Registration | Tens of Billions RMB Cumulative · embodied ⚠️ single-source claim

Registered less than 90 days ago, the company has already landed three consecutive funding rounds, with Hillhouse Capital, HighLight Capital, and Chinaco Ventures each adding on. The company states its valuation now exceeds $1 billion, making it a unicorn — the fastest such run in the embodied-intelligence industry (pace and cumulative amount are company-disclosed figures). Source: 有材 source (WeChat, CN)

Humanoid (UK) | New Round | $150M | Unicorn · humanoid

UK humanoid robotics company Humanoid has closed a $150M round and reached unicorn status, adding another high-valuation player to Europe's humanoid race. Source: 智东西 source

X Square Robot (自变量机器人, Chinese embodied-AI startup) | Four Rounds Raised | Valuation Over $2.8B · embodied

Focused on embodied foundation models (Physical AI foundation model), the company has completed four consecutive funding rounds, with valuation surpassing $2.8B, placing it in the top-valuation tier of China's embodied-foundation-model leaders. Source: AOL.com source

BXI Robotics (半醒具身) | New Round | Tens of Millions RMB · humanoid

A humanoid robot ODM solutions provider; the investor is A-share listed company Suochen Technology, with funds earmarked for next-generation humanoid R&D and overseas market expansion — a listed company using an equity stake to enter the humanoid supply chain. Source: 机器人前线 source (WeChat, CN)

III. Commercialization & Deployment

AgiBot's Genie G2 Max Enters JD's Smart Wolf Warehouse: First Humanoid Robot in a Real Production Warehouse · humanoid

At WAIC on July 17, JD Logistics and AgiBot jointly unveiled the safety-rated, heavy-payload humanoid robot Genie G2 Max, announcing its deployment at JD Logistics' Smart Wolf Warehouse — officially described as the first time a humanoid robot has been deployed in a real warehouse production workflow rather than a trade-show demo. G2 Max is focused on inbound handling and depalletizing: it has human-like arm proportions and a heavy-load force-control system, can grip and move standard totes around the clock, and interfaces with upstream and downstream automation equipment. The Smart Wolf Warehouse is JD's self-developed "goods-to-person" system, with vertical racking over 10 meters tall and storage density about 4 times that of a traditional warehouse, already deployed at scale in Beijing, Wuhan, Chengdu, and the UK. AgiBot's Genie G2 had previously surpassed 15,000 units shipped (previously reported); this move advances "first at-scale deployment scenario" from a shipment figure to embedding in a real production line. JD Logistics says it plans to procure 3 million robots over the next five years. Source: JD Logistics (京东物流JDL) source (WeChat, CN)

UBTECH's Walker U1: Orders Top 13,000 Units Across All Channels, Deliveries Start in September · humanoid ⚠️ single-source claim

UBTECH's "full-size ultra-lifelike humanoid robot" Walker U1, which went on sale in June, had accumulated more than 13,361 units in orders across all channels as of June 30, with deliveries starting September 16. The full-body version is priced at RMB 159,800, with the series spanning RMB 119,800–990,000, and is marketed for companionship (explicitly stated as "not a boyfriend/girlfriend/partner," shipped with custom makeup, not eligible for seven-day no-reason returns). The order volume is notable, but figures come from company disclosure, and delivery and mass production remain to be proven out. Source: 维科网人形机器人 source (WeChat, CN) · 新识研究所 source (WeChat, CN)

Ant Group's Lingbo "Smart Pharmacy" Named a WAIC "Showcase Highlight" · embodied

Ant Group's Lingbo unit deployed a "one brain, many robots" architecture in a smart pharmacy scenario, with multiple robots coordinating on large-scale picking and sorting, earning recognition as a WAIC 2026 "showcase highlight"; described as China's first robotic pharmacy, it was also featured as a conference highlight — a real example of embodied intelligence deployed in a vertical commercial setting. Source: 上观新闻 source

Volkswagen's Moia to Pilot Robotaxi Service in Hamburg · autonomy

Volkswagen's mobility brand Moia has announced a robotaxi pilot launching in Hamburg, Germany, entering local European autonomous-mobility operations — another example of a legacy automaker running its own robotaxi service. Source: 디지털투데이 source

IV. Industry Developments

Xi Jinping Attends WAIC for the First Time, Frames AI as "Moving from the Digital World into the Physical World" · embodied

On July 17, the 2026 World Artificial Intelligence Conference and High-Level Meeting on Global AI Governance opened in Shanghai, with China's President Xi Jinping attending for the first time and delivering the keynote address, stating clearly that "artificial intelligence is moving from the digital world into the physical world." This marks the first time China's top leadership has linked AI's next phase to the "physical world" at the highest-level setting, and industry observers read it as the highest-level policy endorsement yet for embodied intelligence. Earlier signals had been building progressively: in March, the National Development and Reform Commission added "embodied intelligence" to its strategic emerging industries catalog, and in May the Ministry of Science and Technology released a draft Action Plan for Embodied Intelligence Innovation and Development. The speech, combined with the density of industry activity on WAIC's first day — leading companies racing toward capital markets, real factory settings opening up for large-scale validation — still leaves three hurdles unaddressed industry-wide: mass production, real paying use cases, and profitability models have yet to be broadly proven out. A policy green light solves "can it go public," not "can it make money." Source: 天瑞丰年 source (WeChat, CN)

Zoox Issues Software Recall After Thick-Smoke Misjudgment, Covering 105-Vehicle Fleet · autonomy

Amazon's Zoox issued a software update recall covering its entire 105-vehicle fleet after a robotaxi, in June, braked hard and became stuck at a fire scene shrouded in thick smoke with no cones set up to cordon it off — the update adds recognition and response capability for "emergency scenes obscured by thick smoke." No one was in the vehicle at the time and no one was injured; a remote operator ultimately reversed the vehicle out of the scene. This is Zoox's second recall, following a hard-braking issue in March 2025, and the timing is notable — the recall decision was made on July 7, one day before NHTSA's administrator sent a letter to the entire industry demanding robotaxis stop obstructing emergency responders; regulators explicitly characterized "failure to recognize emergency scenes" as a "functional insufficiency" rather than a rare edge case (following on NHTSA's earlier ultimatum, previously reported). Around the same time, San Francisco's mayor called for tighter regulation after a Waymo traffic incident — regulatory scrutiny on U.S. robotaxis is tightening. Source: TechCrunch source

OFweek: Mainland China's Humanoid Robot Production Topped 30,000 Units in H1, Shipments About 23,000 · humanoid ⚠️ institutional estimate

According to OFweek's "2026 Humanoid Robot Industry Insight Report," China produced over 30,000 units of humanoid robots and shipped roughly 23,000 units in H1 2026, with full-year output projected at around 50,000 units. Leading-producer tallies: AgiBot shipped its 15,000th unit in June, with a full-year estimate of about 20,000; Unitree produced roughly 11,000 units of a single model from January–May, with full-year shipments potentially reaching 15,000; UBTECH's Walker U1 has orders exceeding 13,000, deliveries starting in September; and LimX Dynamics' Foshan production line has annual capacity in the tens of thousands. Compared with roughly 16,000 units installed globally in 2025, the production curve is climbing rapidly — these are third-party report estimates. Source: 维科网人形机器人 source (WeChat, CN)

XPeng's MONA L03 Launches Globally in Munich, Starting at RMB 123,800, With Second-Gen VLA · autonomy

Following its second-generation VLA passing local certification in Germany (previously reported), XPeng's MONA L03 launched in Munich starting at RMB 123,800, with plans to reach 65 countries, marketed around "one autonomous-driving stack spanning China and Europe" — requiring only a small amount of local European post-training data to adapt. Chinese autonomous-driving exports are shifting from "simple replication overseas" toward pushing flagship-tier autonomous driving down into mainstream-priced vehicles. Source: 汽车之家 source

NIO Puts Its Automotive-Grade 5nm "Shenji" Chip into Robots, Launches Embodied-Intelligence Development Platform · hardware

NIO showcased at WAIC putting its automotive-grade 5nm chip, "Shenji," to use in robots, and launched an embodied-intelligence development platform, opening its in-vehicle compute and toolchain to robot manufacturers — repurposing its smart-vehicle compute and chip capabilities for the embodied-intelligence space. Source: 驱动之家 source

BrainCo Unveils "World's First" Brain-Controlled Robot Development Platform · adjacent ⚠️ demo claim

BrainCo debuted a brain-computer-interface-driven robot training platform at WAIC, claiming that 10 minutes is enough to achieve "mind control" of a robot (capability based on a live demo and vendor claims). Source: StreetInsider source

MIIT Standards Committee Holds Annual Plenary Session, Building Out Embodied Intelligence's "Rules Foundation" · adjacent

The 2026 annual plenary of the Ministry of Industry and Information Technology's Standardization Technical Committee for Humanoid Robots and Embodied Intelligence, alongside "Standards Week," was held in Shaoxing, Zhejiang, with NavInfo taking the lead as chair unit of the Data Standards Working Group; industry standards for dexterous hands are now open for public comment (previously reported). As the industry races toward mass production, it is simultaneously moving to fill in the foundational rules and data standards. Source: 砍柴网 source

Hardware & Supply Chain

· Dexterous Hand Price Cuts and Volume Ramp: GGII estimates that China's dexterous-hand shipments will surge from 19,200 units in 2025 to 70,200 units in 2026 (+260%), with downstream volume ramp forcing localization of core joint components like micro motors, driving a wave of financing in the micro-motor sector source (WeChat, CN)

· Joyson Electronic (均胜电子): unveiled core robot components including dexterous hands and hybrid solid-liquid batteries at WAIC, marking a Tier 1 auto-parts supplier entering the embodied-intelligence supply chain source

· Zhongda Leader (中大力德): holds roughly RMB 3.2B in orders from Unitree, supplying 63%–80% of its components, yet gross margin fell from 25.77% to 21.44% — "selling more, earning less," reflecting upstream price-war pressure source (WeChat, CN)

FutureX · Physical AI Daily — Issue 60 (07/17)

Shawn — Thu, 16 Jul 2026 14:48:44 +0000

Today's Highlights

· SoftBank exercises its put option, Hyundai buys out Boston Dynamics' last 9.65% stake for $325 million

· Nvidia teams up with the Japanese government to build the world's first national-level physical AI compute base, with 27,500 Rubin GPUs

· RSS 2026 Best Paper FlashSAC: humanoid locomotion sim2real training compressed from hours to minutes

· Harmonic reducer prices halve in three years, bulk quotes down to just over 2,000 yuan

· NHTSA to finalize new driverless-vehicle rules by 2028, replacing hardware-design clauses with a "behavioral capability" test

I. Paper Progress

FlashSAC: compressing humanoid locomotion sim2real training from hours to minutes · locomotion

RSS 2026 Best Paper. The conference this year accepted 210 of 708 submissions, a 29.7% acceptance rate, with 8 papers making the finalist list. The work tackles an old problem in off-policy reinforcement learning for high-dimensional robot control — slow, unstable training. It goes against convention by reducing the number of gradient updates while simultaneously scaling up model size, data throughput, and the experience replay buffer, then controlling value-network error accumulation by bounding weight, feature, and gradient norms. It outperforms PPO and various off-policy baselines overall across 10 simulators and 60-plus tasks, and the most visible result is a humanoid locomotion sim2real experiment where training time shrank from hours to minutes. The authors are from Holiday Robotics, KAIST, KRAFTON, TU Darmstadt, and KTH — Holiday Robotics is the same company that set a new Korean Series A record of KRW 155 billion just last week.

Donghu Kim, Youngdo Lee et al. (Holiday Robotics × KAIST × KRAFTON) · RSS 2026 Best Paper · Analysis: QbitAI source (WeChat, CN)

NeuralActuator: giving low-cost robot arms a sense of how much force they're feeling, without torque sensors · perception

RSS 2026 Best Systems Paper. Friction, backlash, and temperature drift throw off the traditional current-to-torque relationship — a major source of sim2real gap for cheap robot arms. This work uses a transformer combined with differentiable simulation to jointly predict actuator torque, external contact force, contact probability, and motor state, letting robots sense applied force at deployment without dedicated force/torque sensors. It was validated on three classes of robot arms ranging from about $500 to over $30,000, and the team also released the NAD dataset collected via bimanual teleoperation.

Zhiyang Dou et al. (MIT CSAIL) · RSS 2026 Best Systems Paper · Analysis: QbitAI source (WeChat, CN)

DAPL: robots learn to "use the environment as leverage," extracting objects without grasping · manipulation

Selected for the RSS 2026 finalist list. The framework first uses a world model to learn the dynamics of collisions, sliding, and tipping among objects in cluttered scenes, then uses these dynamics representations to guide reinforcement learning — so the robot no longer relies solely on direct grasping, and can also complete "extrinsic dexterous manipulation" like pushing, nudging, and flipping by making contact with surrounding objects and the environment. In unseen simulated cluttered scenes, success rate improved by more than 25 percentage points over baselines such as grasping and teleoperation, and reached about 50% across 10 categories of real-world scenes; it has already been used for shelf-picking tasks. The authors emphasize these behaviors emerged autonomously from the model rather than being tuned via reward hacking.

Yixin Zheng, Jiangran Lyu et al. (CASIA × BAAI × Galbot × Peking University × SJTU) · RSS 2026 Final List · Analysis: QbitAI source (WeChat, CN)

UniTac: letting VLA "imagine" the feel of touch before contact · manipulation

The difficulty in bringing touch into VLA isn't a lack of data, but that the data is locked inside heterogeneous sensors — the same object can produce completely different signals across different tactile sensors. UniTac splits touch into two separately modeled parts — "object physical properties" and "sensor configuration" — and learns a unified representation from over 400,000 publicly available tactile videos, then predicts latent tactile states from vision and feeds them into VLA, shifting touch from post-contact feedback to a pre-contact prior. On the understanding side it scores 66.51 on PHYSICLEAR-Test, ahead of models like Octopi; on the generation side it achieves an average SSIM of 0.836 across four sensor types. The value lies in scenarios like paper cups or towels — objects that "break if you squeeze wrong" — where correcting after contact is already too late.

UniX AI × Zhejiang University × MIT × Oxford × Yale × SJTU · Accepted at ECCV 2026 · Analysis: QbitAI source (WeChat, CN)

GigaWorld-Policy-0.5: a world-action model that no longer generates video at inference time · world-model

21↑ on Hugging Face trending. Existing world-action models (WAMs) generally generate explicit future video at inference time, and the compute overhead is heavy enough to break real-time closed-loop control. This version takes an "action-centric" approach — using dense supervision from future visual dynamics during training, decoding only actions at inference, and introducing a Mixture-of-Transformers to split visual dynamics modeling and action generation across different experts, achieving 85ms inference latency on a single RTX 4090. The team also used an agentic AutoResearch pipeline to automatically search training configurations, cutting out a large amount of manual tuning.

GigaWorld Team, Angen Ye et al. · arXiv 2607.13960 source

WANDA: synthesizing a trainable world from a single real-robot demonstration · benchmark

The labor cost of teleoperation and UMI-based data collection simply doesn't scale for open-world mobile manipulation. WANDA's approach is to squeeze maximum value from a single demonstration: it first reconstructs a background Gaussian splat and robot-object interaction trajectories from RGBD observations as a "world substrate," then rearranges contact-rich interaction segments across many spatial configurations and strings them into new trajectories via whole-body motion planning, synthesizing realistic observations across diverse 3D worlds generated from everyday photos. Experiments show policies trained this way simultaneously gain long-horizon robustness, spatial generalization, and cross-environment generalization, and naturally support cross-embodiment data generation — it has already been zero-shot deployed on another mobile manipulation robot with a different form factor.

Lingxiao Guo, Huanyu Li, Guanya Shi · arXiv 2607.13154 source

APT-RL: quadrupeds autonomously switch gaits for high-speed traversal in the wild, using only onboard sensors · locomotion

The difficulty in multi-skill locomotion lies in the handoff between skills — when to switch gaits and to which one is usually decided by a hand-crafted state machine. This framework uses an action-pretrained transformer to unify multi-skill locomotion, enabling autonomous skill switching and high-speed traversal across complex terrain using only onboard perception and compute, validated from rugged wilderness to urban environments.

Jun-Gill Kang, Jaehyun Park et al. · arXiv 2607.13579 source

Other papers today: M4World, a multi-view multimodal driving world model supporting object-level interaction and minute-scale streaming generation (arXiv 2607.14005 source); Kepler-Encoder-v0.1, treating proprioceptive state as a modality fused with vision and force/torque into a shared latent space — vision-only features reading force have an R² as low as 0.10 (arXiv 2607.13522 source); Anchor-Align, using a frozen VLM copy for layer-wise distillation to prevent behavior-cloning fine-tuning from overwriting pretrained representations (arXiv 2607.13429 source); Semantic Anchoring, anchoring action representations back to a semantic manifold and splitting them into shared-semantic and private channels (arXiv 2607.13597 source); Reverse to Advance, exploiting a task's directional asymmetry, using an "easy reverse task" to lower teleoperation collection cost for hard tasks (arXiv 2607.13455 source); PhysClaw-0, letting language corrections persist and be reused across turns so supervision cost scales with the number of problem types rather than conversation length (arXiv 2607.14047 source); UniPhysGen, automatically converting raw 3D assets into simulation-ready assets with articulation semantics and physical properties (arXiv 2607.13586 source); From Pixels to States, examining interactive world models through four game-engine dimensions, with a data engine featuring over 90 hours of "Black Myth: Wukong" data (arXiv 2607.14076 source); Grounded world models, proposing five reference points for embodied world models drawn from biological neural circuits (arXiv 2607.13560 source); showing the SIGReg objective is equivalent to variational free energy, providing JEPA world models with a normative active-inference interpretation (arXiv 2607.13612 source).

Open Source · Tools · Benchmarks

· X-TouchMind V1 and TacVerse 1k: Qianjue Robotics released what it says is the industry's first tactile-focused VTLA embodied foundation model, paired with a 1,000-hour native visual-tactile dataset — 100% tactile data coverage, with vision, touch, force, pose, and high-frequency dynamics captured together rather than tactile data added on after the fact to vision-only data; the model uses a System 0-2 hierarchy, with System 0 dedicated to high-frequency tactile correction source (WeChat, CN)

· RLinf v0.3: an update to the embodied reinforcement learning framework jointly built by Infinigence AI and Tsinghua University, with five capability upgrades spanning model ecosystem to real-robot deployment source

· GPUSimBench: a benchmark for GPU-accelerated simulators establishing standards for scalability, physical consistency, and computational determinism, pointing out that unchecked scaling sacrifices reproducibility (arXiv 2607.13059 source)

· Industrial Dexterity Benchmark: an integrated hardware-software benchmark platform for industrial dexterous manipulation, covering cable routing, connector insertion, and precision assembly — tasks still done by hand today (arXiv 2607.14021 source)

· HRIBench: writes collaborative tasks as executable scripts with explicitly modeled roles, temporal dependencies, and coordination constraints, filling a near-blank spot in VLA evaluation for human-robot interaction structure (arXiv 2607.13056 source)

II. Funding and Deals

Boston Dynamics | Equity Buyout | ~$325 Million | 9.65% Stake · humanoid

Following yesterday's reports that Hyundai planned to buy out SoftBank's holding, the trigger and price of the deal were confirmed today: SoftBank actively exercised its put option, selling its remaining 9.65% stake back to Hyundai Motor Group for about $325 million, giving Hyundai 100% ownership. The price is worth a closer look — when Hyundai bought 80% in 2021 for 1 trillion won (about $674 million), Boston Dynamics was valued at about 1.24 trillion won; that valuation has now surpassed 30 trillion won. At the current valuation, 9.65% should be worth roughly $2 billion, meaning SoftBank exited at a discount under the option terms agreed years ago. Hyundai also laid out a timeline for Atlas: entering its Metaplant America factory in Georgia for parts sequencing in 2028, expanding to component assembly by 2030. Multiple Korean media outlets also noted that consolidating ownership clears the way for a Boston Dynamics IPO in the US.Source: Yonhap source | Bloomberg source

Senior Auto (Chinese autonomous trucking startup) | Series C | RMB 300 Million · autonomy

Jointly invested by Xingzheng Capital and Yidao Capital, the funds are primarily for R&D on next-generation production-grade autonomous driving solutions. Founded in 2020, this driverless heavy-truck company follows a "closed → semi-open → open" progressive rollout, monetizing through operating-as-a-service and technology licensing, and has already achieved scaled commercialization in ports and closed parks. Its product is centered on world-model-driven decision-making, covering onboard intelligence, fleet management, and cloud dispatch. Around the Yiyongzhou corridor, it has connected a complete chain across three node types — Ningbo-Zhoushan Port, the Yiwu Suxi rail yard, and the Yongjin Expressway — spanning port drayage, yard transfer, and trunk-line transport. This approach of "using cash flow from closed scenarios to sustain the push into open scenarios" is one of the few autonomous heavy-truck business models in China currently proven to work.Source: PEdaily source

microagi (Germany) | Seed Round | $55 Million · embodied

Led by Hummingbird, with participation from Northzone, LocalGlobe, Village Global, and redalpine — the largest seed round in German startup history. Founder Bercan Kilic left a job as an aerodynamics engineer at Red Bull Racing shortly after landing it in 2023, and started the company last year out of a hacker house in Munich. The company neither builds robots nor foundation models; instead it records workers doing their jobs using cameras and sensor gloves, then uses that data to fine-tune off-the-shelf robot models into specialists for a customer's production line. Five companies currently contribute data on its platform, with one preparing for deployment. Its consumer-facing offshoot, shift, went viral earlier this year by offering "free housework in New York in exchange for first-person recordings," and this week started offering free private chef services in San Francisco.Source: Business Insider source

Think Surgical (US) | Debt Financing | Up to $65 Million · adjacent

The orthopedic surgical robotics company reached an agreement with healthcare investment firm Symbiotic Capital worth roughly RMB 440 million, to advance commercialization of the handheld micro-orthopedic robot TMINI. Choosing debt over equity typically signals the company places a higher priority on avoiding dilution and has confidence in predictable existing cash flow.Source: MedRobot source (WeChat, CN)

SwitchOn (India) | New Round | $8 Million · industrial

A physical AI company focused on manufacturing visual quality inspection, funding will go toward scaling production-line inspection deployments.Source: Inc42 source

III. Commercialization and Deployment

Machina Labs wins Lockheed Martin JASSM missile qualification contract · industrial

This marks the first time components produced by a Machina factory have advanced to qualification testing for a defense missile system — a higher bar than a general "partnership intent." The company's proprietary RoboCraftsman cells autonomously handle loading, metal forming, scanning, trimming, and drilling, turning digital designs directly into production-grade metal parts; the delivered component is a fuel tank metal structure, produced at its Los Angeles factory using its RoboForming process combined with precision laser welding, and handed to Lockheed Martin for testing and system integration. Follow-on production volume will be handled by the under-construction Factory 3: 200,000 square feet, capable of housing up to 50 RoboCraftsman cells. The key bottleneck for robots serving critical structural parts in defense supply chains is replacing tooling and dies with seven-axis flexible forming — once this qualification hurdle is cleared, the marginal cost of switching to a new part becomes a different order of magnitude from traditional stamping.Source: The Robot Report source

Tesla Robotaxi registered vehicles in Texas rise to 175, up 58 in a single day · autonomy

According to a Texas autonomous-vehicle tracking website, Tesla's registered autonomous vehicles in the state have reached 175, with 58 added in a single day — a marked acceleration in pace. An important caveat: this is the registered count, not the actual number in active operation — Electrek reporters previously conducted an on-the-ground count across thousands of square miles of the Austin metro area and found the number of vehicles actually operating far below the registered total. On the other hand, one data point favors Tesla: according to NHTSA incident reports, its Austin fleet has not reported a single at-fault accident since mid-April this year; the only new incident was a Model Y that was rear-ended while parked.Source: Autohome source | LeiPhone source (WeChat, CN)

Stryker's Mako orthopedic surgical robot power system goes fully commercial in the US, cumulative surgeries surpass 2.5 million · adjacent

The system, aimed at knee replacement, has moved from a limited release to full US commercial launch. The same day, the company disclosed that its Mako platform has completed over 2.5 million surgeries across 47 countries — a solid data point that surgical robotics has already cleared the commercialization-validation bar, and also the real benchmark that general-purpose humanoid research results — like papers on Unitree's G1 performing surgery — should actually be measured against.Source: MassDevice source | Stock Titan source

DHL Malaysia uses AI drones for warehouse inventory checks, cutting cycle time by 92% · industrial

Drones automatically perform cycle counts inside the warehouse, cutting inventory-check time by 92%. Warehouse counting is a textbook case of "high-frequency, low-skill, extremely hard to staff" work, and one of the easiest categories for automation ROI to pencil out clearly.Source: BusinessToday Malaysia source

Ukraine's Nurse TB20 ground robot enters mass production · adjacent

This ground unmanned platform has begun mass production, mainly for battlefield casualty evacuation and resupply. Wartime demand is compressing the "prototype to mass production" cycle for ground robots well outside the normal defense-industry pace.Source: UA.NEWS source

IV. Industry Developments

Nvidia and the Japanese government build the world's first national-level physical AI compute base · hardware

Nvidia announced a partnership with Noetra to build a Vera Rubin AI factory, configured with 13,750 Vera CPUs and 27,500 Rubin GPUs, 140 megawatts of data center capacity, using Vera Rubin NVL72 racks and the DSX platform, interconnected via Spectrum-X Ethernet. This compute will support Japan's METI-launched FRONTia project (developing multimodal foundation models for AI robotics and physical AI), feeding Japanese manufacturing know-how and real industrial data into the models; the pretrained weights Noetra trains will be opened up to Japan-based model developers and companies, alongside open models such as Nemotron, Cosmos, and Isaac GR00T. Announced the same day on the industry side: Japanese manufacturing giants Fanuc, Yaskawa Electric, Kawasaki Heavy Industries, and Toyota's Woven City said they will build their respective physical AI capabilities on NVIDIA Cosmos. Jensen Huang put it bluntly: "Japan invented modern manufacturing, and now it is building the AI factories that will power the next industrial revolution." Set alongside this week's Hyundai buyout of Boston Dynamics and Samsung's plan to build a humanoid base, the three East Asian manufacturing powers are each betting on physical AI according to their own asset structures: Japan is betting on infrastructure and open models, Korea on complete units and bringing embodiment in-house.Source: NVIDIA Newsroom source | Chosun Ilbo source

Nvidia launches Jetson Thor T3000/T2000, pushing robot compute into mainstream price points · hardware

The T3000 delivers 865 FP4 TFLOPS of compute at roughly half the size and power draw of the T5000, with a Blackwell GPU, an eight-core Neoverse Arm CPU, 32GB of LPDDR5X, and 273GB/s of bandwidth; despite the halved footprint, inference performance on multimodal workloads such as large language models, VLMs, VLA, and world foundation models remains comparable to the T5000. The entry-level T2000 offers 400 FP4 TFLOPS and 16GB. Nvidia pointed to a very concrete motivation right now: moving to the T3000 cuts costs amid rising memory prices — and newly released Jetson agent skills can automatically optimize memory usage, letting developers drop a memory SKU tier within the same product class without losing performance. The customer list already includes 1X, Agile Robots, Amazon Robotics, Boston Dynamics, Fanuc, Hitachi, Techman Robot, and UBTech. This extends Nvidia's edge platform lineup from 70 TOPS all the way up to 2,000 TFLOPS.Source: NVIDIA Blog source | VideoCardz source

WAIC unveils its "top ten exhibits," only one humanoid robot makes the list · humanoid

The 2026 World Artificial Intelligence Conference opens tomorrow, and its top-ten-exhibits list includes only one humanoid robot — Zhiyuan Robotics' (Chinese humanoid startup) full-size commercial-scenario model Yuanzheng A3 Ultra. Another related selection is Ant Group's Lingbo smart pharmacy, which emphasizes deploying robots without any store renovation under a "one brain, many machines" approach. The shift in judging criteria is more notable than the list itself: doing backflips no longer counts — what counts is running stably on production lines and in stores. Also from the show floor: the number of embodied-intelligence exhibitors rose from about 80 last year to over 200 this year, and roughly 60 full-size humanoid robots will be deployed at the venue for guiding and answering questions.Source: Cailian Press source | Purple Cow News source

Samsung Electronics reportedly plans KRW 13 trillion humanoid robot base in Gumi · humanoid

According to Korean outlet The Elec, the base would be overseen by Samsung Electronics' DX division, with the humanoid production line tentatively sited at an unused factory in Gumi's No. 1 Industrial Complex, while Samsung SDS builds a supporting AI data center locally. ⚠️ Reported by Korean media, unconfirmed Samsung has not publicly confirmed this, and both scale and timeline remain to be finalized. Samsung's existing positioning offers context: it has raised its stake in Korean robotics company Rainbow Robotics to 35% through investment and exercising call options, participated in a RMB 1 billion strategic round for Xingdong Jiyuan (Chinese humanoid startup) in March, and this week Samsung Ventures joined Walden Robotics' $300 million seed round.Source: WeChat Humanoid Robotics source (WeChat, CN)

Public consultation launched on industry standard for humanoid robot dexterous hands · embodied

The standard sets out a full set of technical requirements and standardized test methods for dexterous hands, covering degrees of freedom in appearance, motion performance, perception capability, grasping and manipulation capability, common device interfaces, electromagnetic compatibility, environmental adaptability, and operational reliability, along with accompanying test procedures — filling a regulatory gap in this segment. Dexterous hands are currently the highest-priced single component and also the one with the least cross-comparable specifications — every manufacturer quotes its own degrees of freedom, payload, and tactile channels differently. Once the standard takes effect, buyers will at least have a common yardstick.Source: Household Service Robot Committee source (WeChat, CN)

China adds 27 vocational-education majors including "humanoid robot engineering technology" · humanoid

China's Ministry of Education added 27 new majors to its vocational education catalog, with humanoid robot engineering technology among them. What actually bottlenecks factory-floor deployment at the ten-thousand-unit scale is often not the model, but people who can commission, maintain, and script the workflows — a change to the major catalog signals the supply side is beginning to gear up on an industrialization timeline rather than a research one.Source: China News Service source

Nvidia reshuffles autonomous-driving leadership, Wu Xinzhou (Nvidia autonomous driving VP) consolidates authority · autonomy

According to an exclusive from 36Kr, Nvidia carried out a major personnel reshuffle in its autonomous driving business, with authority consolidating around Wu Xinzhou. This follows yesterday's reported hire of a former Qualcomm autonomous-driving VP — the convergence of "ex-Qualcomm" and "ex-Xpeng" factions within Nvidia's automotive unit is becoming clearer.Source: 36Kr source (WeChat, CN)

Hardware · Supply Chain

· D-Robotics Sunrise S600: this high-compute chip aimed at embodied intelligence has entered mass-production validation, endorsed by over 20 leading embodied-intelligence companies (carried over from yesterday's report) source (WeChat, CN)

· Xinjian Transmission: a planetary roller screw manufacturer pushing for a ChiNext IPO, planning to raise RMB 2.822 billion to expand production; the company was the first to achieve large-scale mass production of planetary roller screws in China in 2022, and is a Tier-1 screw supplier for Tesla Optimus source (WeChat, CN)

· Jinghua New Material: its subsidiary Jingzhigan has established deep business collaboration with multiple dexterous-hand and embodied-intelligence companies source

· SK On: solid-state batteries for robots must first prove their economics — the battery maker struck a rare note of caution about this new customer segment, humanoid robots source

V. This Week's Observations

NHTSA to finish new driverless-vehicle rules by 2028: shifting from "what's in the car" to "what the car can do" · autonomy

NHTSA Administrator Jonathan Morrison confirmed to Bloomberg that the Trump administration plans to complete new autonomous-driving safety requirements by the end of this term, i.e. by 2028. The direction is to replace decades-old design-based rules that assume "a person sits behind the steering wheel" with performance-based standards: soliciting public and industry input to define what "behavioral capabilities" a self-driving car should have, then designing objective tests to determine compliance — for example, whether it can safely navigate around a double-parked fire truck, rather than what control hardware is installed in the vehicle. Morrison's framing centers on predictability: "manufacturers producing vehicles must be able to know for certain whether their vehicles meet the standard." Hardware-side deregulation is already underway: last month a proposal to amend FMVSS No. 135 to remove the brake pedal requirement for fully self-driving designs was put forward, mandatory steering wheels are also under review, and clauses on rearview mirrors, gear displays, wipers, sun visors, defrosting systems, and tire labels have all been proposed for removal — directly benefiting purpose-built robotaxis like Cybercab and Zoox. On the other side, regulatory pressure hasn't eased: Waymo, the largest robotaxi operator in the US (about 4,000 vehicles), recently recalled nearly its entire fleet and paused highway operations after vehicles drove through construction zones at highway speed; the NTSB and others have investigated multiple incidents of robotaxis passing stopped school buses; a January incident where a vehicle struck a child near an elementary school remains under investigation.Source: Automotive World source | Bloomberg source

Morgan Stanley: 2026 is the inflection year for robotaxis, a trillion-dollar market by 2040 · autonomy

Morgan Stanley's latest global robotaxi outlook judges the industry to be moving from technical trials to commercial scale-up, with a timeline of: mainstream operators reaching per-mile breakeven by 2028, scaled operations by 2030, a global fleet of 2.5 million vehicles by 2035, and a trillion-dollar-scale market by 2040. Cost deflation is the biggest catalyst — lidar and compute chip costs are falling 3-5% annually, and per-mile operating costs in some Chinese cities have already dropped below $0.50; the report cites two manufacturer-specific figures: Tesla's Cybercab targeting a cost of $0.30/mile, and Chinese vehicle BOM costs compressed to $35,000-40,000. Regionally, the report expects the US and China to lead (with Waymo and Tesla projected to split nearly 70% of US autonomous-driving miles by 2032), while Europe lags 3-5 years behind due to fragmented regulation slowing profit realization, and Japan and Korea treat it more as a solution to labor shortages. ⚠️ Single-bank view This diverges from Citi's assessment last week that "scaling up is a decade-long marathon" — the disagreement centers on the timing of breakeven.Source: Morgan Stanley research source (WeChat, CN)

TrendForce: companion humanoid robot output value only $1.1 billion by 2030 · humanoid

TrendForce forecasts that caregiving and emotional-companionship demand will push the companion humanoid robot market to $1.1 billion by 2030. The significance of this figure is that it puts a yardstick on the "humanoids pivoting to consumers" narrative — $1.1 billion is a rounding error next to Barclays' forecast last week that the humanoid market could reach $200 billion in a decade. This week UBTech was read as "sidestepping the labor-task logic, pivoting toward emotion and companionship, and going after consumers," but per this forecast, companionship alone won't support a leading manufacturer's core business, at least not before 2030.Source: TrendForce source

This week's supply chain roundup: harmonic reducer capacity has come online, and prices have collapsed · hardware

This is the most notable supply-chain signal this week, and it runs counter to the "supply chain chokepoint" narrative — it's not a shortage, it's a glut. On the capacity side: global harmonic reducer production capacity was 4.897 million units in 2024, jumping straight to 6 million units in 2025; against that, GGII data shows global industrial-robot harmonic reducer consumption was only 1.2187 million units in 2024 (795,500 in China) — existing capacity can absorb all global industrial-robot demand with plenty to spare. Prices have collapsed accordingly: the average price of a standard Chinese-made unit was about RMB 4,500 in 2023 and RMB 3,200 in 2024; by 2026, quotes of just over RMB 2,000 for bulk orders are common, with some smaller models falling below RMB 1,000 — a halving in three years. Three forces are compounding: after a single factory's annual output exceeds 300,000 units, economies of scale push unit cost down more than 30%; new entrants supply near cost price to secure test reports from benchmark customers (one company quoted 40% below the average price just to earn "designated supplier" status); and downstream demand has shifted from industrial robots to humanoids and collaborative robots, which are extremely cost-sensitive — cutting RMB 1,000 off a single reducer saves over RMB 10,000 on a full unit. But there's a hard floor to how far prices can fall: material and processing costs for a standard unit run about RMB 1,500; last year a manufacturer in southern China bought reducers priced at RMB 1,600, and within three months of deployment saw a 30% return rate. Leading makers are responding by digesting existing capacity rather than expanding — Leaderdrive's capacity utilization rose from 42.67% in 2024 to 67.76% in 2025, with output jumping from 251,700 to 433,700 units, while it pushed back a planned 1-million-unit capacity expansion project from the end of 2026 to the end of 2028; even so, its 2025 revenue of RMB 571 million rose 47% year-over-year, with net profit up 121% — prices are falling, yet profit growth is outrunning revenue growth.Source: Zhiti Xinju source (WeChat, CN) | Daoyi Institute source (WeChat, CN)

Reconciling demand: humanoid unit sales went from 50 to 15,000 in two years, reducers sell 430,000 units a year · hardware

Following on from the previous item, here's the demand-side accounting. Per MIR DATABANK, sales of complete-form humanoid robots in China were about 50 units in 2023 and about 15,000 units in 2025; a single standard humanoid uses more than 14 reducers, corresponding to third-party reducer sales of about 430,000 units for China's humanoid sector in 2025. The prevailing configuration is "harmonic drives for the upper body, planetary for the legs" — harmonic drives are dominated by Leaderdrive, Laifual, Radial Drive, and Tongchuan, while planetary drives are led by Zhongda Leader and NUODAR, which together account for over 70% of that market and primarily supply Zhiyuan Robotics and Unitree; cycloidal reducers sit between the two in load capacity and precision — theoretically a better fit for humanoids — but have complex manufacturing and higher costs, and remain in small-batch validation. Worth noting is a loosening at the component end: in the past only a handful of manufacturers worldwide could make 6mm micro harmonic drives, with overseas units selling for over RMB 2,000 each; Chinese manufacturers achieved mass-production breakthroughs this year and drove costs down to under RMB 1,000, recalculating the materials-cost math for dexterous hands. By contrast, another part of the chain remains tight: a humanoid uses 14 planetary roller screws, accounting for about 19% of a unit's total materials cost — more than harmonic drives and servo motors combined — and depends on imported high-end grinding machines with long validation cycles, which is why Xinjian Transmission is pushing for an IPO this week to raise RMB 2.822 billion for expansion. Reducer oversupply and roller-screw scarcity coexisting — that's the real shape of this supply chain right now.Source: MIR Rui Industry source (WeChat, CN) | Robot Chain source (WeChat, CN)

FutureX · Physical AI Daily — Issue 59 (07/16)

Shawn — Wed, 15 Jul 2026 14:45:23 +0000

Today's Highlights

· Xiaomi open-sources 38-billion-parameter world foundation model U0, topping the World Arena leaderboard for embodied video generation

· Walden exits stealth with $300 million backed by Toyota and Nvidia, $1.1 billion valuation

· RoboParty closes two rounds totaling nearly RMB 500 million, with CATL leading exclusively

· XPeng's second-generation VLA passes German certification without retraining, sprinting toward Europe deployment

· LimX Dynamics releases COSA 0.5 humanoid brain system, argues "the model is not the brain"

I. Research Papers

Xiaomi U0: Turning a world foundation model directly into an embodied synthesis engine · world-model

Following yesterday's preview of "world models as data factories," Xiaomi officially open-sourced the model and released the paper today. U0 is a 38-billion-parameter multimodal autoregressive model that unifies text-to-image generation, image editing, embodied scene generation, cross-embodiment transfer, and embodied video generation into a single framework via unified next-token prediction, initialized from the open-source EMU3.5 (Qwen-3-32B). It is the first to support high-quality multi-view scene generation across multiple robot embodiments, ranks first on the World Arena benchmark for embodied video generation, and lifts the generalist policy π₀.₅'s out-of-distribution success rate on real-world manipulation tasks from 36.9% to 63.2%; paired with anti-diagonal decoding, 1024² image generation is roughly 83x faster than native autoregressive generation. The key takeaway is that it validates the idea that a foundation world model can serve both as an embodied world model and as a scalable data engine.

Xiaomi Robotics · arXiv 2607.11643 source · Analysis: Shuyuan AI source (WeChat, CN)

FlowWAM: Using optical flow as a "universal action language" to unify world action models · world-model

Uses optical flow as a unified action representation for world action models (WAM), aligning well with pretrained video generators while carrying enough motion cues for control—an attempt to answer "how to find an action representation for video generators that is both alignable and drivable."

Yixiang Chen et al. · arXiv 2607.13017 source

VistaVLA: Adding an explicit 3D scene representation to VLA · vla

Existing VLAs map language and 2D images directly to actions, lacking scene-level 3D representations and struggling to reason about spatial layout and geometric constraints. VistaVLA uses 3D Gaussians to ground geometry and semantics, enabling manipulation policies to explicitly reason about spatial relationships.

Mohan Liu et al. · arXiv 2607.12356 source

TrustVLA: Defending VLA's "visual backdoors" at inference time · vla

End users can't audit a VLA's deployment pipeline; a poisoned model can behave normally on clean footage yet hijack long-horizon policies the moment a small visual trigger appears, before any failure surfaces. This work proposes a mechanism-guided inference-time defense that identifies and suppresses backdoor triggers without retraining—a safety angle rarely seen in VLA work but a critical one.

Pinhan Fu et al. · arXiv 2607.12571 source

Jetson-PI: Getting VLA to run real-time control on Jetson · vla

VLA is computationally heavy, causing high latency and low control frequency when deployed on low-power boards like the Jetson Orin. This work uses "lookahead-aligned asynchronous inference" to enable onboard real-time control, addressing a real bottleneck for on-device deployment on physical robots.

Zebin Yang et al. · arXiv 2607.12659 source

GaitSpan: A humanoid that can walk shouldn't have to learn to run from scratch · locomotion

Humanoid robots shouldn't have to relearn locomotion from zero just to jog or run. GaitSpan lets gait "grow" from walking to running, avoiding the skill fragmentation that comes from existing methods relying on preset gait timing or imitating motion clips.

Kwan-Yee Lin et al. · arXiv 2607.12114 source

Other papers today: ChunkFlow eliminates boundary jitter in chunked action heads (arXiv 2607.12992 source); Reducing Temporal Redundancy cuts VLA inference time redundancy for a speedup (arXiv 2607.12287 source); ExToken uses structured exploration to improve VLA reinforcement fine-tuning efficiency (arXiv 2607.12931 source); DenseReward learns dense rewards from synthesized failures (arXiv 2607.13033 source); Robust In-Hand Manipulation achieves robust in-hand manipulation without external perception (arXiv 2607.12105 source); Hy-Embodied-VLM-1.0, Tencent's efficient embodied VLM from its Hunyuan team (arXiv 2607.12894 source); MAMMOTH is a robust off-road end-to-end policy under missing modalities (arXiv 2607.12965 source); TerraZero is a zero-demonstration self-play driving simulator (arXiv 2607.13028 source).

Open Source · Tools · Benchmarks

· WorldArena 2.0 Challenge: a "real exam" for world models launched by Jiqizhixin (Chinese AI media outlet) and others; Track 2 places world models inside an online reinforcement-learning loop—receiving policy actions, predicting the next observation, and returning a reward—testing whether a model can "serve as a training environment" rather than just generate a video clip source (WeChat, CN)

· Ruka-v2: an open-source tendon-driven dexterous hand under $1,500, with an upgraded wrist and e-flesh soft fingertips, capable of writing, flipping pages, and holding a pen source (WeChat, CN)

II. Funding and Deals

Walden Robotics | First round out of stealth | $300 million | $1.1 billion valuation · industrial

Co-led by Toyota and Deviation Capital, with participation from Nvidia, Boeing, Samsung, Prologis, and CoreWeave, among others. The company spun out of the Toyota Research Institute (TRI) this past January; its founder and CEO is Russ Tedrake, an MIT professor and former head of TRI's large behavior models effort, pursuing a "full-stack physical AI" approach that builds on Diffusion Policy and Large Behavior Models. The most notable point isn't the valuation: its robots have been in production at a Toyota North America factory since February, going from pilot to production line in under two months, targeting labor-constrained industries including automotive, aerospace, semiconductors, electronics, logistics, and life sciences.Source: The AI Insider source

RoboParty | Angel++ round + Pre-A round | Nearly RMB 500 million · embodied

CATL led the Pre-A round exclusively, with Shunwei Capital, Matrix Partners China, and Xiaomi's strategic investment arm adding on; the two rounds together total nearly RMB 500 million. The company builds an open-source full-stack embodied platform, with a founding team of 22-year-olds from the Harbin Institute of Technology background; CATL leading the round in person signals that battery and industrial capital are pushing upstream into embodied platforms along the "hardware-components" chain.Source: Embodied Emergence (Chinese embodied-AI media) source (WeChat, CN)

Monumental (Netherlands) | New round | $32 million · industrial

Led by Khosla Ventures, founded by two former Palantir employees; the funds will go toward expanding its fleet of 150 brick-laying robots across Europe and into the United States, addressing labor shortages in construction trades in places like the UK.Source: Tech Funding News source

Cytronic | Seed round | $13.5 million · industrial

Reported exclusively by Axios; builds warehouse fulfillment (order picking) robots, targeting the mid-sized warehouse segment that large-scale automation hasn't yet reached.Source: Axios source

Yuejiang Technology | Upcoming ChiNext IPO review | First "H-to-A" listing in the Greater Bay Area · embodied

Yuejiang, a maker of collaborative and humanoid robots, goes before the review committee next week and will become the first company in the Guangdong-Hong Kong-Macao Greater Bay Area to complete an "H-to-A" listing (moving from a Hong Kong listing to an A-share listing).Source: 21jingji (Chinese financial media) source

Deyi Medical | Series A++ round | Tens of millions of RMB · embodied

A fund under Tiantu Capital invested exclusively; the company builds AI-plus-healthcare embodied robots and has now closed two funding rounds within the year.Source: Tiantu Capital source (WeChat, CN)

III. Commercialization and Deployment

Agibot's Genie G2 hits 15,000 units cumulative, launches G2 Max with JD Logistics · humanoid ⚠️ vendor claim

According to Agibot (Zhiyuan Robotics, Chinese humanoid startup), its Genie G2 reached its 15,000th unit off the line by late June and has been deployed on production lines at a Longqi Technology factory, taking roughly three months to go from 10,000 to 15,000 units; for reference, Omdia data put total global humanoid shipments last year at around 13,000 units, meaning this single batch already exceeds last year's global total—Agibot held nearly 40% of global humanoid market share last year. Its paired "Data Collection 2.0" dual-track collection (real robot plus simulation) aims to shore up sim-to-real gaps, and is also central to its partnership with Aishida—though it should be noted that Aishida assembles the Cooper quadruped, not the full humanoid unit. Today, JD Logistics and Agibot jointly launched the embodied robot Genie G2 Max. Leading in production volume is not the same as a "technical leap"—the mass-production numbers are, more than anything, favorable conditions for accumulating data.Source: Zhizhen Guan (Chinese robotics media) source (WeChat, CN)

Unitree's G1 goes abroad to "lay bricks": Japanese construction firm puts a humanoid on-site · humanoid

A Japanese construction company has deployed Unitree's G1 humanoid robot on a construction site—another instance of a Chinese humanoid robot going overseas into a real industrial setting, rather than a stage demonstration.Source: IT Home (Chinese tech media) source

Gausium named world's No. 1 commercial cleaning robot supplier by IDC · industrial

An IDC report ranks Gausium as the world's top commercial cleaning robot supplier by shipments; the company says it is shifting from winning flagship contracts toward scalable growth. Cleaning remains one of the few commercial service robot categories that has already achieved scale.Source: PR Newswire source

Boston Dynamics trials robot dog Spot for door-to-door package delivery · adjacent ⚠️ pilot

Boston Dynamics demonstrated its Spot robot completing "the last fifty feet" of door-to-door package delivery; this remains a pilot demonstration, not scaled delivery.Source: The Verge source

Warehouse and retail robots land at multiple sites: Amazon, Tesco, Colruyt · industrial

Amazon has reportedly been building a roughly $1 billion robotic warehouse in western Sydney; UK grocer Tesco is piloting a 6.5-foot-tall robot that patrols stores to automatically detect shelf out-of-stocks; Belgium's Colruyt has teamed up with KION to establish an AI warehouse robotics center. Warehousing and shelf fulfillment are becoming among the fastest-scaling deployment scenarios.Source: The Urban Developer source

IV. Industry Developments

LimX Dynamics releases COSA 0.5 humanoid brain system, reiterates "the model is not the brain" · humanoid

LimX Dynamics released a teleoperation-free, single-take demo of its humanoid brain system LimX COSA 0.5: the robot Oli continuously performs long-horizon household chores including hanging laundry, putting things away, stacking and moving boxes, and deep bending to pick up objects. Founder Zhang Wei states plainly that "the model is a skill, the system is the brain"—the brain should be an operating system that orchestrates VLM/LLM/VLA and other models and tools, not a single ever-growing model. The system has three layers: S2 cognitive scheduling (roughly 1 Hz), S1 skill layer (VLA/world action models generating whole-body motion policies), and S0 whole-body motor control (LimX WBT, roughly ten million parameters, running at 1000 Hz on the physical body). LimX reports WBT's average joint position error at 12.85 millimeters, slightly better than SONIC, the industry's strongest whole-body tracking model it benchmarks against (13.75 millimeters). This contrasts with Figure's emphasis on an end-to-end unified neural system, putting the debate over "is the brain a system or a large model" on the table; the single-take demo remains a capability showcase—scaling will depend on whether it can be reliably reproduced.Source: DeepTech (Chinese tech media) source (WeChat, CN)

Stardust Intelligence releases embodied foundation model Lumo-2 and physical AI agent Philia · world-model ⚠️ vendor claim

Lumo-2 is positioned as a "household Latent World-Action Model": before acting, it predicts task-relevant physical changes in a compressed latent space, then generates actions accordingly—sitting between traditional VLA and full video world models, and covering 22 household tasks such as catching a ball, flipping food in a pan, weighing rice, and tying a bow. According to its paper, end-to-end latency on a single RTX 5090 drops from about 253.66 ms to 93.53 ms, outperforming baselines such as π0.5 and Fast-WAM overall. Also released alongside it, Philia turns robots into "physical executors" that agents can call, exposing standardized capabilities via a Robot Gateway, pointing toward a "systems engineering" phase for embodied AI.Source: AI Faner (Chinese tech media) source (WeChat, CN)

XPeng's second-generation VLA passes localization certification in Germany, He Xiaopeng attends in person · autonomy ⚠️ certification claim

XPeng's Turing AI-assisted-driving NGP second-generation VLA model has completed localization certification testing in Europe, with He Xiaopeng traveling to Germany personally for the certification; the company states the same model adapts to European road signs and traffic regulations without retraining, aiming to be the first Chinese brand to deploy advanced driver-assistance in Europe. This is the company's own certification claim; commercial deployment still awaits local regulatory approval.Source: IT Home (Chinese tech media) source

Hyundai reportedly moving to buy out SoftBank's stake in Boston Dynamics · humanoid

According to Korean media reports, as a deadline approaches, Hyundai Motor is reportedly moving to acquire SoftBank's stake in Boston Dynamics, further consolidating ownership of the veteran robotics company.Source: Korea Financial News source

UBTech signs "the Mercedes-Benz of dealerships" group, humanoid robots enter auto dealership channels · humanoid

UBTech has signed a sales partnership for its humanoid robots with a dealership group known as "the Mercedes-Benz of 4S stores," which becomes its first authorized partner in the Beijing-Tianjin-Hebei region—robots are beginning to reach customers through auto dealership channels.Source: Sina Finance source

Policy and capital both ramp up ahead of WAIC, embodied AI gets its own track · adjacent

This year's World Artificial Intelligence Conference (WAIC 2026) is establishing a dedicated embodied AI zone for the first time, splitting "AI computing" and "embodied intelligence" into two independent core tracks; concurrently, Guangdong launched a RMB 100 billion strategic emerging industry fund, and robotics ETFs have seen consecutive days of inflows. On the official data side, China's National Bureau of Statistics reports 3D printing equipment output up 48.5% year-over-year in the first half of the year. Policy, capital, and industry are moving in sync, setting the tone for the sector right now.Source: Cailianshe (Chinese financial media) source

Former Qualcomm autonomous-driving VP set to join Nvidia, "Qualcomm alumni" and "XPeng alumni" converge · autonomy

Qualcomm's former VP of engineering for autonomous driving is set to join Nvidia's autonomous driving team, seen as another reinforcement of Nvidia's automotive autonomous driving effort.Source: Sina Finance source

Hardware · Supply Chain

· Galaxea (Chinese robotics startup): will showcase at WAIC a dexterous-hand solution priced at the RMB 10,000-level (⚠️ vendor claim, conference promotion) source

· Linker Hand: claims peak monthly output of over 4,000 units and over 80% global share of high-degrees-of-freedom dexterous hands, with a valuation of RMB 40 billion (⚠️ vendor claim) source (WeChat, CN)

· Cost breakdown (Yiou, Chinese tech media): a 17-actuator, 5-sensor linkage-driven dexterous hand has a BOM cost of about RMB 51,800, with actuation and transmission accounting for roughly 60% of it—this is the reference point against which the "RMB 10,000-level" pricing claim should be read source (WeChat, CN)

· Horizon Robotics' Sunrise S600: has entered mass-production validation for its embodied AI chip, claiming recognition from 20+ leading customers (⚠️ mass-production validation claim) source

· Joyson Electronics: will showcase a full robotics component lineup at WAIC—dexterous hands, head assemblies, and embodied AI "brains" source

FutureX · Physical AI Daily — Issue 58 (07/15)

Shawn — Tue, 14 Jul 2026 14:45:44 +0000

Today's Highlights

· After 4 months of on-the-job training in a car factory, Xiaomi's humanoid robot's nut-fastening success rate has risen to 98%, just one percentage point short of the human benchmark

· Amap released and open-sourced the general-purpose world model workshop ABot-World Studio, which can run locally on a single RTX 5090 with continuous inference for over an hour

· The HiPhi (Huawei-backed EV brand) G9 became the first vehicle in Beijing to receive an L3 autonomous driving road test license rated for 120 km/h

· Two tactile/dexterous-hand deals in one day: Xense Robotics raised over 100 million yuan, and DexRobot's Series A of several hundred million yuan was led by Shanghai Electric

· Xiaomi's U0 treats the world model as a data factory, lifting pi0.5's out-of-distribution success rate from 36.9% to 63.2%

I. Papers

Xiaomi-Robotics-U0: World models don't belong in the control loop — use them as a robot data factory instead · world-model

This paper offers a pragmatic answer to "how should world models actually be used in robotics": rather than doing real-time planning for the robot at inference time, it "manufactures" scenarios the robot has never seen before training. Xiaomi (Chinese consumer electronics and EV company) proposes a unified embodied synthesis framework that uses a world foundation model to generate multi-view-consistent, geometrically coherent, embodiment-constrained interaction data — directly filling the most expensive gap in real-robot data: out-of-distribution scenarios. The result: pi0.5's out-of-distribution task success rate rose from 36.9% to 63.2%, nearly doubling, with no additional real-robot data collection.

Xinghang Li et al. (Xiaomi) · arXiv 2607.11643 source · Commentary: ICCV source (WeChat, CN)

ABot-N1: Turning vision-language navigation into a general-purpose foundation model · autonomy

The most-discussed paper of the day in the community (HF↑76). Most existing navigation foundation models are monolithic policies that map observations directly to actions, which are prone to coordinate drift and struggle with long-tail semantics. ABot-N1 separates deep spatial reasoning from multi-task generality and then unifies them, aiming for a VLN backbone that generalizes across embodiments and scenarios. It belongs to the same ABot series that Amap (Alibaba's mapping subsidiary) released together today, alongside the two papers below.

Ruiyan Gong et al. (Amap) · arXiv 2607.10383 source

ABot-3DWorld 0: Turning a sentence, image, or video into an explorable 3D world · world-model

This is the 3D generation engine behind Amap's ABot-World Studio, which launched today — the model has been open-sourced. Its core is a compact representation called the Spatial Generative Primitive (SGP) — a high-quality panorama paired with a spatial point cloud — used to uniformly describe arbitrary 3D spaces. This converges text, image, and video inputs into a single, real-time-explorable world representation, directly serving robot training and simulation data collection.

Mingchao Sun et al. (Amap) · arXiv 2607.11673 source

ABot-AgentOS: Giving robots an "operating system" with lifelong memory · benchmark

HF↑66. VLM/VLA models have advanced perception and action prediction, but long-horizon tasks still lack a general runtime layer — someone still has to manage reasoning, memory, tool calling, result verification, and cross-embodiment execution. This paper abstracts that layer into a general-purpose robot Agent OS with cross-modal lifelong memory — addressing a gap in "upper-layer orchestration" that papers tend to overlook but that shows up daily in real-world deployment.

Jiayi Tian et al. (Amap) · arXiv 2607.10350 source

EgoSteer: Feeding dexterous hands "steerability" from first-person human video · vla

Steerability — doing what you're told, and changing course when told to — is a capability general-purpose policies should have but dexterous-hand systems largely lack. The bottleneck is data: there's a shortage of large-scale, language-aligned, action-precise demonstrations. EgoSteer offers a full-stack solution that uses first-person human video to scale up pretraining for dexterous-hand VLA models, followed by post-training with a very small amount of real-robot data.·

Yifan Zhong et al. · arXiv 2607.09701 source

Mixture of Frames Policy: Action denoising shouldn't happen in just one coordinate frame · manipulation

Manipulation is inherently multi-frame: fine end-effector motion is simplest to express in the tool frame, while transport, righting, and whole-body coordination make more sense in the base frame. This paper has the policy denoise separately in multiple reference frames and then blend the results, targeting bimanual mobile manipulation — a structural correction to diffusion/flow policy action representations rather than just more parameters.

Dian Wang et al. (Stanford) · arXiv 2607.11884 source

VIA: Letting foundation models control robots by "looking at" and "clicking on" an interface · vla

General foundation models already have strong vision and reasoning; their weak point is closed-loop control. Sadigh's group's approach isn't to hard-train an end-to-end policy, but to give the foundation model a visual interaction interface as a middle layer, letting it operate the robot agent-style — realigning "what the model is good at" with "what the robot needs."

Hengyuan Hu et al. (Stanford) · arXiv 2607.11119 source

Cycle-World: Using "backward prediction" to curb drift in long-horizon video world models · world-model

In autoregressive diffusion for long video generation, errors accumulate step by step into structural collapse — a wall every team hoping to use video world models as simulators runs into. This paper adds a backward-prediction cycle-consistency constraint, letting the model check for itself whether "the present can be recovered from the future," thereby suppressing long-horizon generation drift.

Zihan Su et al. · arXiv 2607.11836 source

Other papers today: Towards Human-level Dexterous Teleoperation, a teleoperation system for in-hand regrasping and finger gaiting (arXiv 2607.11481 source); WALA, learning executable latent actions simultaneously from action-labeled demonstrations and action-free video (arXiv 2607.11397 source); WarpMPC, solving a hundred thousand MPC subproblems in parallel on a GPU (arXiv 2607.11603 source); TAC-LOCO, integrating touch into quadruped whole-body control for dynamic loco-manipulation (arXiv 2607.10132 source); A Minimalist Retargeting-Guided RL Recipe, testing whether the "retargeting + RL" recipe from humanoid whole-body control transfers to dexterous manipulation (arXiv 2607.11874 source); From World Action Models to Embodied Brains, a roadmap from world-action models toward open-world physical intelligence (arXiv 2607.11689 source).

Open Source · Tools · Benchmarks

· Horizon Robotics simulation-training generation engine: open-sourced; the same scene can run across MuJoCo, Genesis, and IsaacSim simultaneously — the first time robot training can reuse a single scene asset across simulators source (WeChat, CN)

· ABot-World series: alongside ABot-World Studio, Amap open-sourced two model weights — ABot-World0 (interactive video generation) and ABot-3DWorld0 (3D scene generation) source (WeChat, CN)

· GigaWorld-1: a cost-reduction approach for real-robot policy evaluation, using generative world models to replace part of real-robot trial runs source (WeChat, CN)

II. Funding & Deals

Xense Robotics (千觉机器人) | New round | Over 100 million yuan · hardware

Strategic investment from a top embodied-intelligence industry player and Chigo (Chinese home appliance maker), with follow-on from Tianji Capital. Founded in May 2024 by Ma Daolin, an associate professor at Shanghai Jiao Tong University and the sole global winner of the ICRA 2021 Best Paper Award, the company builds an integrated solution combining multimodal tactile sensors with vision-tactile data and models, and says it has served over 300 leading industry clients. The funding will go toward industrializing high-precision, three-color-light vision-tactile sensors. Touch sensing is moving from "a modality in papers" to a component with real orders and a place in the supply chain — this is the second tactile-sensing company to raise over 100 million yuan this week.Source: Sensor Expert Network source (WeChat, CN)

DexRobot (灵巧智能) | Series A | Several hundred million yuan · embodied

Investment from Shanghai Electric. The company specializes in robot end-effectors and dexterous manipulation systems, describing its mission as solving the "last centimeter" of a robot's interaction with the physical world. The funding will go toward product iteration for its dexterous hands and scaled deployment across multiple industries. Industrial capital (rather than pure financial VC) betting directly on end-effectors points to real industrial demand for "hands that can do fine work."Source: Robotics Open Community source (WeChat, CN)

LimX Dynamics (逐际动力) | Pre-IPO round | Nearly $200 million | Post-money valuation 15 billion yuan · humanoid

Following disclosure of this round earlier in the week, the company officially confirmed closing and finalized the valuation today. Investors include IDG Capital, Lens Technology, Hefei Binhu Financial Holdings Group, GGG Group and Redstone VC, Huashan Capital, and Legend Capital, among others; the company says it has raised $400 million cumulatively over the past six months. For context: Crunchbase data shows global robotics startup funding reached $18.8 billion in the first half of 2026, already surpassing the full-year 2025 total of $15 billion.Source: Shanghai Securities News source (WeChat, CN)

Quadric | Series C | $46 million · hardware

Builds a programmable AI chip platform for edge inference. Robot compute solutions still rely heavily on general-purpose GPUs, and programmable edge chips are one path around that dependency.Source: Pulse 2.0 source

NovaTech (South Korea) | New round | 7 billion won (~$5 million) · adjacent

With participation from Hyundai Motor Group. The company builds robot orchestration software, coordinating multiple robot brands via its PiPER platform; this round's funding will go toward expanding into the North American market. An automaker betting on the "orchestration layer" rather than the robot itself is a side signal that multi-robot coordination on production lines is a real pain point.Source: Chosunbiz source

Ondas acquires DZYNE | $875.8 million · adjacent

Unmanned systems company Ondas acquired defense drone maker DZYNE, expanding its defense capabilities. This is the largest robotics-related acquisition of the week by value.Source: The Robot Report source

III. Commercialization & Deployment

Xiaomi's humanoid robot completes 4 months of "internship" in a car factory, nut-fastening success rate hits 98% · humanoid

Lei Jun (Xiaomi founder) disclosed the latest data on the humanoid robot's on-the-job training at the factory: at the self-tapping nut installation station, the bilateral operation success rate rose from 90.2% to 98% over 4 months of iteration — just one percentage point short of the human-operator pass rate. More notably, newly unlocked stations — installing the center console side panel and folding/recycling parts bins — both reached 90% success rates; the side panel is large, irregularly shaped, and highly flexible, with the task involving multiple pick-and-place cycles at multiple positions, marking the first time a robot has completed long-duration handling of a flexible workpiece on a real production line. What makes this significant isn't the 98% figure itself, but that it's one of the rare publicly disclosed, time-series production-line training curves: over the past year, the industry has mostly offered single demos or "X hours without human intervention," whereas "same station, four months, success rate climbing from 90.2 to 98" provides a previously almost never-disclosed metric — the rate of iteration. Of course, 98% is still a single-station training result, still some distance from production-line metrics like overall cycle time, uptime, and defect cost.Source: QbitAI (Chinese tech media) source (WeChat, CN)

DEEP Robotics' quadruped robot deployed at Switzerland's Leibstadt nuclear power plant · industrial

DEEP Robotics (云深处) announced its quadruped robot has been deployed at Switzerland's Leibstadt nuclear power plant, handling digitized inspection and maintenance — replacing human workers on instrument readings, thermal imaging, and vibration baseline collection in radiation zones. Nuclear power is one of the few scenarios where quadruped robots have already proven a viable paying-customer logic (humans shouldn't linger, paths are fixed, data is structurable), and it's a relatively realistic entry point for Chinese robot makers into European industrial clients.Source: ACCESS Newswire source

Tesla expands Robotaxi to Miami, develops wheelchair-accessible model in Texas · autonomy

Tesla's Robotaxi service launched in Miami this week, and the company confirmed it is developing a wheelchair-accessible autonomous vehicle model in Texas. The accessible model isn't just a product-line addition: many U.S. jurisdictions require ride-hailing operators to maintain a minimum proportion of accessible vehicles in their fleets, and Robotaxi can't truly replace taxi fleets without clearing this bar.Source: Electrek source

Turkey's Tripy Mobility to deploy Robotaxis using WeRide-Vision (Chinese autonomous driving supplier) tech · autonomy

Chinese autonomous driving solutions provider WeRide-Vision (天瞳威视) has won the Robotaxi project for Turkish mobility platform Tripy Mobility. Chinese autonomous driving solutions are moving beyond exporting vehicles or licensing algorithms toward the deeper commitment of supplying full-stack solutions to local operators.Source: Sina Finance source

Tesco trials 5-foot shelf-inspection robot · industrial

UK retail giant Tesco has begun trialing a 5-foot-tall aisle-inspection robot that automatically identifies out-of-stock shelf locations. Retail inspection is one of the few mobile-robot scenarios that can create value through perception alone, without grasping — but it remains at the trial stage, with no large-scale procurement yet.Source: The Grocer source

Serbia to launch China-backed humanoid robot production in August · humanoid

Serbia will launch a humanoid robot production line with Chinese backing in August. ⚠️ Plan stage Production scale and the specific robot model have not yet been disclosed; the outcome remains to be verified.Source: IntelliNews source

IV. Industry News

Amap releases general-purpose world model workshop ABot-World Studio, all models open-sourced · world-model

Alibaba's Amap officially launched ABot-World Studio today and opened it for testing, while also open-sourcing the two models behind it: ABot-World0 (interactive video generation) and ABot-3DWorld0 (3D scene generation). The product's technical framing is clear: world models have previously split into two separate paths — interactive video generation offers strong immersion but interaction ends once the video finishes, while 3DGS scene generation offers high fidelity but isolated, non-interconnected worlds. ABot-World Studio brings both paths into a single platform: users input text or an image and get a real-time-interactive, shareable world, which can be saved either as video or as a 3DGS file. It also includes a "portal" mechanism — passing through a door jumps you into another complete 3D world, weaving previously isolated scenes into an infinitely explorable network. Two engineering figures are worth industry attention: the company says it can be deployed locally on a single RTX 5090, and a single continuous inference session runs stably for over 1 hour without crashing or quality degradation — whereas comparable world models today typically generate for around 1 minute. For embodied-AI teams, this means the barrier to simulation-based data collection has dropped from a GPU cluster to a single card.Source: Amap source (WeChat, CN)

HiPhi G9 approved for Beijing L3 road test license, first model rated for 120 km/h · autonomy

The HarmonyOS Intelligent Mobility HiPhi G9 has officially received a Beijing municipal L3 autonomous driving road test license, becoming the first vehicle model in Beijing to qualify for L3 testing at 120 km/h. Its technical approach uses multi-sensor fusion combining LiDAR, vision, and millimeter-wave radar, with a Huawei 896-line LiDAR unit standard across the lineup; it has passed simulation testing, closed-course testing, and four rounds of validation covering functional safety, safety of the intended function, cybersecurity, and data security. Its L3 operational design domain covers nighttime scenarios and supports lane changes on highways/expressways. Viewed against last week's publicized mandatory national standard for autonomous driving (effective 2027, requiring heterogeneous redundant sensing for L3), the fusion-sensing approach's first-mover advantage in China's L3 approval process is now being converted into an actual license.Source: Yicai source

ByteDance's head of generative AI, Yuan Zehuan, departs to start a world-model venture · world-model

Yuan Zehuan, who previously led ByteDance's generative AI direction, has left to start a venture in the world-model space. ⚠️ Single source Combined with this week's report that ByteDance's Seed world-model team is exploring autonomous driving (officially denied as having any business plan), world models are simultaneously drawing away both capital and talent from big tech companies.Source: AIcore (Chinese tech media) source (WeChat, CN)

Beijing builds and opens embodied-AI robot pilot validation platform for the power industry · industrial

An embodied-AI robot pilot validation platform for the power industry has been built in Beijing and is now officially open for external service. Pilot platforms address what robot makers most lack — batch validation under real power-grid conditions (high voltage, live wires, narrow corridors) rather than benchmark testing in a lab. Such industry pilot facilities, led by grid operators and power research institutes, are becoming the actual chokepoint through which embodied AI must pass on its way from demo to industry admission.Source: Guandian.cn source

JD's robotics ecosystem adds two more partnerships: Huilun Technology, Luming Robotics · adjacent

Following the groundbreaking of its Guangzhou RoboBase, JD (Chinese e-commerce giant) signed strategic partnerships this week with embodied-AI companies Huilun Technology and Luming Robotics; Luming disclosed a partnership target of 1 billion yuan in sales over 3 years, with multiple robot hardware products set to launch on JD in 2026. JD's approach differs from both robot makers and algorithm companies: trading channel and supply-chain access for ecosystem position, positioning itself as the sales and infrastructure backbone for robotics.Source: Gasgoo source

14 institutions across Japan, the US, and Europe join forces on physical AI research; Bengio joins Noetra · adjacent

Led by Japan, 14 institutions across the US, Europe, and Japan have formed a physical AI research collaboration network aimed at building a Japan-native physical AI model; Yoshua Bengio has joined the affiliated company Noetra to help drive its commercialization. Japan's underlying motivation here is easy to read: its robot hardware and components remain world-class, but its "brain" layer — model capability — now clearly lags behind China and the US.Source: The Japan News source

South Korea's SKT and KT launch AI-RAN trial networks to lay groundwork for physical AI services · adjacent

South Korea's two largest carriers, SK Telecom and KT, are deploying "AI base stations" at industrial sites, launching an AI-RAN pilot worth roughly $11.5 million as a core piece of South Korea's "AI highway" initiative. Carriers are treating physical AI as the next growth driver after 5G/6G — robots on factory floors need to upload video and receive policy updates in real time, and the network layer is indeed an often-overlooked bottleneck.Source: Seoul Economic Daily source

Hardware · Supply Chain

· Dedicated dexterous-hand chips: CAS Wireless Semiconductor (中科无线半导体) has launched China's first integrated dexterous-hand chip solution, the CT-2001H/CT-1906H, paired with a CT-HS01 4D spectral sensor achieving 0.1mm positioning precision and microsecond-level latency; the company says it cuts hardware costs by 30% and is natively compatible with VLA models source (WeChat, CN)

· Boya Intelligence (伯牙智能) Gaoshan S1: a dexterous hand weighing just 500 grams with 30N of fingertip force, developed by a team co-founded by a former Alibaba executive and a robotics professor source (WeChat, CN)

· Yuequan Bionics (月泉仿生) Yingshou Y-Hand M1: the full hand weighs 299.7 grams; rather than competing on degrees-of-freedom specs, it uses a bionic tension-compression transmission design source (WeChat, CN)