Today's Highlights
· NHTSA opens an investigation hours after Cybercab hits the road, probing Tesla's basis for self-certifying compliance
· Draft revision of China's Road Traffic Safety Law adds a dedicated autonomous driving chapter, putting liability for violations while the function is activated on automakers
· Huandong Technology (RV reducer maker, subsidiary of Shuanghuan Transmission) withdraws its STAR Market IPO at the last moment, taking China's top RV reducer maker off the capital markets for now
· UBTECH lands a 1.5 billion yuan project in Shanxi, sending humanoid robots underground into coal mines
· Wandercraft signs 12 Calvin-40 customers, with Renault planning to deploy 350 units within 18 months
· Axis open-sources 50,000 Franka simulation trajectories, lifting π0.5's score on LIBERO-Plus by 4.9 points
· Stardust Intelligence's SmoothRL raises real-robot dynamic throwing success rate from 39% to 94%
Research Progress
1,500 hours of bimanual household demonstration data open-sourced, used to train VLA model XR-2 · vla
Public data for bimanual household manipulation has long been stuck below a thousand hours; this release provides 1,500 hours of demonstrations in one go, covering everyday household tasks, and uses this corpus to train the VLA model XR-2. The authors run scaling experiments along two axes: adding more expert demonstrations, and using DAgger-style correction data generated through real-time human intervention for post-training. Success rate climbs steadily with data volume along both axes, with no inflection point observed within the range explored. The dataset is released alongside the paper.
Jiafeng Xu et al. · arXiv 2609.03591 source
BRIDGE: putting humanoid morphology design and whole-body control into a single optimization framework · locomotion
Designing hardware separately from whole-body control is a longstanding root cause of humanoid robots "moving unlike humans." This paper proposes a data-driven framework for co-optimizing morphology and control, along with a new metric that jointly measures kinematic retargeting fidelity to human motion and dynamic tracking performance. Against three baseline humanoids — Bumi, K1, and ToddlerBot — it achieves SOTA on every metric. The result is an open-source 88cm-tall platform called Bridge, with the control policy released alongside it.
Jianren Wang et al. · arXiv 2609.03497 source
MINERVA: 0.54M parameters is enough to hit 95.1% on LIBERO · benchmark
The LIBERO leaderboard is now dominated by billion-parameter-class VLAs; this paper instead asks how much capacity the benchmark itself actually requires. A policy with just 0.54M parameters runs 2,000 rollouts across four standard suites, achieving an average success rate of 95.1% — only 2.4 points below the π0.5 result reported by LeRobot, with 7,700 times fewer parameters and only 5–9 milliseconds per chunk on a laptop CPU. A more notable finding: a task-ID permutation probe that only alters the task-index mapping drops success rate to near-random, suggesting that standard LIBERO's instruction conditioning is largely memorizing tasks. On the LIBERO-Plus perturbation set, performance slides to 46%–56%.
Kohei Sendai et al. · arXiv 2609.03715 source
FailBench: using VLMs as judges of robot task success or failure — even the best scores only 0.77 · benchmark
Using VLMs to judge whether a robot task succeeded is becoming the default approach, yet cross-domain reliability has rarely been measured. FailBench collects 2,197 manipulation attempts from 14 public sources, 75% of whose failures occurred naturally; testing 13 VLM detectors, the best achieves an average balanced accuracy of only 0.77. Models specifically fine-tuned for failure detection actually score consistently lower than general-purpose VLMs and their own pretrained baselines. On contact-rich assembly tasks, scores fall below 0.60, and error analysis further finds that when evidence is ambiguous, models systematically bias toward judging "success."
Zaruhi Navasardyan et al. · arXiv 2609.03611 source
EGR: teaching VLAs which sensor channels actually matter · vla
Trained on limited and homogeneous demonstrations, VLAs readily learn spurious correlations between sensor modalities — a failure mode the authors call modality entanglement. EGR uses a frame-by-frame, sensor-by-sensor task-relevance signal for gating, adds two consistency constraints during training, and incurs zero extra cost at inference. In real-world tests with a dual-arm Kinova plus physical distractors, success rate rose from 30% to 85%; on a single-arm platform combining vision with GelSight tactile sensing, it rose from 55% to 70%.
Yue Yang et al. · arXiv 2609.03142 source
WISE: scheduling a world model's imagination cuts GPU time by 80% · world-model
World models let policies try and fail in imagination, but this is compute-expensive, and rollouts that run too long accumulate prediction error and feed back unreliable supervision signals. WISE's approach is to schedule imagination, triggering it only in states relevant to interaction, running bounded multi-view rollouts, and evaluating candidate futures using progress and completion signals. It improves both π0 and π0.5, cutting GPU compute time by roughly 80% compared to full-scale imagination.
Chenhao Zhang et al. · arXiv 2609.03681 source
RoboTok: turning internet human videos into searchable robot supervision · manipulation
Robot data is expensive, and long-tail tasks are especially hard to cover. Given a query video of human manipulation, RoboTok retrieves relevant demonstrations from internet video to train dexterous-hand policies. The method learns an implicit motion space from 3D hand trajectories in a human-centric reference frame, making manipulation behaviors comparable across different camera views, scenes, and occlusions, while remaining compact enough to support continuous internet-scale indexing. Both retrieval relevance and downstream task success rate outperform existing robot data retrieval methods.
Howard Qian et al. · arXiv 2609.03199 source
Raw data may stay local, but household information still leaks through what's exported · perception
Home robots keep raw sensor data on-device, but the structured representations exported to planners, cloud services, and logs can still leak household information. The authors quantify this across 120 AI2-THOR scenes: three navigation-export representations all achieve a task success rate of 1.000 and an average path ratio of 0.898, yet their linkability at the representation layer ranges all the way from 0.532 to 0.970. Replacing explicit goal labels with goal regions drops target-category macro-F1 from 1.000 to 0.077 while success rate remains at 0.995; geometric coarsening, meanwhile, comes at the cost of collision-detection utility. Neither dropping fields nor adding abstraction yields a general-purpose privacy ranking — privacy must be evaluated task by task.
Yuqiao Xu, Erman Ayday · arXiv 2609.03055 source
Other papers today: FWBC-VLA adds contact awareness to wheeled-legged robot VLAs via sensorless residual torque estimation, validated on whiteboard wiping and door opening (arXiv 2609.03889 source); R2S-Eval replaces repeated real-robot trials with real-to-sim calibration plus VLM preference scoring (arXiv 2609.03276 source); SV-WAM builds a surround-view world-action model into end-to-end driving while controlling inference overhead (arXiv 2609.03602 source); Drive-HWM guides driving decisions using dynamic latent variables from a hierarchical world model (arXiv 2609.03572 source); LaPla connects VLM discrete reasoning to continuous control via latent-space alignment planning (arXiv 2609.04070 source); MulDP enables quadrupeds to autonomously parkour-navigate complex terrain (arXiv 2609.03984 source); one paper re-examines world models in safety-critical embodied systems, arguing that high prediction likelihood and visual fidelity do not guarantee that evidence needed for safe decisions has been preserved (arXiv 2609.03774 source); and another offers a unified survey of robot learning spanning representation learning, VLA, and world models (arXiv 2609.03927 source).
Open Source · Tools · Benchmarks
· Axis Sim Dataset V1: a Franka simulation manipulation dataset open-sourced by Axis Robotics, comprising over 50,000 human teleoperation trajectories, 207 manipulation tasks, and over 60,000 scene variants, using a simulated Franka Research 3 as the embodiment. Continued pretraining of π0.5 on this data raised LIBERO-Plus success rate from 83.9% to 88.8%, outperforming a comparably sized RoboCasa365 baseline by 37.3%; the company says performance kept improving as data scale grew from 25% to 100%. The dataset was co-built with researchers from UC Berkeley, Johns Hopkins, and the University of Michigan, with V2 planned to expand to 1.2 million trajectories and 1,200 tasks. source
· HandEdit: a dataset and evaluation benchmark open-sourced jointly by AgiBot (Chinese embodied-AI robotics company) with Shanghai Jiao Tong University and Fudan University, for the task of replacing human hands in first-person video with a specified robot embodiment while preserving object state, contact relationships, and background. It spans over 200 million image-editing samples and over 300,000 video clips, drawing on five egocentric datasets — EgoDex, ARCTIC, OakInk2, HOI4D, and HO-Cap — covering 26 URDF embodiments, over 600 scenes, and over 1,100 objects. The benchmark uniformly tests 11 editing methods including GPT-Image, Nano Banana, FLUX, and Qwen-Image-Edit, and one finding is that high VLM-assigned scores do not track with whether the interaction is actually correct. source
· SmoothRL: an online reinforcement learning framework released by Stardust Intelligence (Chinese robotics startup)'s foundation model team, with a technical report published on its website. Large-model inference is slow, and real-world deployment generally relies on asynchronous inference, so the actions a model "planned" and the actions the robot "actually executed" end up mismatched; SmoothRL splits an action chunk into committed, executing, and discarded regions based on execution state, computing gradients only over the executing region. Using a π0.5 policy fine-tuned per task on the cable-driven Stardust S1 embodiment, dynamic throwing success rate rose from 39% to 94% (250 rollout episodes), placing a pen cap rose from 8% to 83%, and unboxing a package rose from 30% to 90%. source
Funding and Deals
Huandong Technology (RV reducer subsidiary of Shuanghuan Transmission) | STAR Market IPO | Spinoff terminated, all filings withdrawn · hardware
On the evening of September 3, Shuanghuan Transmission announced it was terminating the spinoff listing of Huandong Technology and withdrawing all application documents, citing "significant changes in market conditions since the plan was first drawn up." Counting from the application's acceptance in November 2024, the push toward listing lasted under ten months. Huandong makes RV reducers, the core component in the major hip, shoulder, and knee joints of industrial robots; its share of the Chinese market reached 24.98% in 2024, up from just 10.11% four years earlier, over the same period biting Nabtesco's Chinese market share down from 51.77% to just over 40%. The sharpest question across two rounds of regulatory inquiry concerned customer concentration: sales share from the top five customers rose from 79.45% to 92.12% during the reporting period, with Estun, Efort, Inovance, and Siasun accounting for nearly all of its revenue, while its largest customer, Estun, swung to a loss of over 800 million yuan in the same period. The company's answer was that it had cut prices to gain share — gross margin fell from 42.47% to 35.36% in 2024, in exchange for market share jumping from 18.89% to 24.98%; in 2025, revenue was 438 million yuan, non-GAAP net profit was 75 million yuan, and gross margin recovered to 36.57%. The withdrawal doesn't change demand-side fundamentals: of the 10 to 14 joints in a humanoid robot, four to six heavy-load joints depend on RV reducers.Source: Phoenix Finance source
XTEND | SPAC merger listing (NYSE: XTND) | $1.5 billion valuation · adjacent
Israeli defense robotics company XTEND completed its merger with JFB Construction Holdings, renaming itself XTEND AI Robotics and beginning trading on the NYSE on September 4, with the merger valuing XTEND at $1.5 billion. The company was founded in 2018 by CEO Aviv Shapira along with three other co-founders, building human-operated, machine-executed tactical drone systems. Its list of strategic investors includes Eric Trump, Protego Ventures, Unusual Machines, and Tel Aviv University's TAU Ventures.Source: Globes
Weifen Zhifei (Hangzhou) | Series A2 | Hundreds of millions of yuan · adjacent
Led by Puhua Capital, with Hongxin Fund, Wuyuan Capital, Shenzhen Capital Group, Hongtai Fund, and BV Baidu Ventures participating. Founded in 2024 and led by Gao Fei, a doctoral supervisor at Zhejiang University, the company works on embodied aerial intelligence, aiming to let aircraft navigate autonomously and complete tasks in unknown environments with no GPS, no communications, and no lighting. Mining is where it has been most proven in practice — its P300 product replaced what would have been over two years of work at a large Chinese mining company in six months, covering more than 300 mining sites; it is now shipping in small batches with hundreds of units on order. Reporting on the funding round differs: JW Insights records it as Series A2, while Sohu's same-day funding roundup records it as Series A+.Source: JW Insights source
Hivebotics (Singapore) | Series A | $6 million · adjacent
Led by Vertex Ventures Southeast Asia and India, with participation from Fareast Land Development and bathroom fixture manufacturer Rigel. Its robot, Abluo, is fitted with a multi-jointed robotic arm to clean toilets, urinals, sinks, and floors in commercial restrooms, addressing what the company says existing commercial cleaning robots can only do for floors. Over the past 12 months, Abluo has logged roughly 10,000 operating hours across 20 sites in Asia, Europe, the Middle East, and North America, including hospitals, airport terminals, and shopping malls, with the company saying it replaces 30 minutes of manual labor with a five-minute human inspection. Co-founder and CEO Rishab Patwari said the company started with restrooms because they are the hardest setting — wet, cramped, and full of moving parts.Source: The Business Times source
Shenpu Intelligence | Pre-A+ round | Hundreds of millions of yuan · embodied
Announced complete on September 3. Funded by Kexi Capital, an intellectual property fund under Shanghai Science and Technology Innovation, Goldenpool Investment, Bozhou Industrial Investment, Chengshang Capital, and Qianrong Capital, with existing investor Linear Capital following on. The company positions itself as an AI-native embodied intelligence systems provider, building its system around four pillars: models, data, embodiment, and scenarios.Source: Guandian.cn source
Commercialization and Deployment
Cybercab begins carrying passengers in Austin, NHTSA opens investigation hours later · autonomy
NHTSA announced an investigation into Tesla on Friday morning, just hours after the first Cybercabs hit the streets of Austin. The vehicle has no steering wheel, no pedals, and no conventional rearview mirrors — none of the manual controls required under Federal Motor Vehicle Safety Standards remain. Tesla told regulators that, as is customary, it self-certified the vehicle's compliance with all FMVSS. NHTSA's filing states it is investigating "the processes and technical data Tesla relied on when certifying the Cybercab," as well as the extent to which the company determined that certain federal standards do not apply to this vehicle. NHTSA Administrator Jonathan Morrison left room for both sides: "NHTSA fully supports the safe development and deployment of autonomous vehicles. But as the federal regulator, we need to ensure all laws are being followed." Zoox went down this same road before — it self-certified compliance in 2022, was hit with a special order the following year triggering the same kind of audit inquiry, saw its commercialization timeline slowed, and eventually completed a Part 555 temporary exemption process, receiving final approval only in July 2026. Fleet size is another variable: according to Texas Department of Motor Vehicles data, Tesla newly registered 38 Cybercabs for its local robotaxi fleet as of August 31, bringing the total to 45 units. The vehicle relies on 8 HD cameras paired with an end-to-end neural network, with no lidar and no high-precision maps, using the same AI4 compute platform found in the Model 3 and Model Y. Tesla says that at scale, cost per mile could fall to $0.20. In China, plans remain limited to static display — starting mid-September, the vehicle will be shown in Beijing, Shanghai, and other cities, with no sales involved and no indication of plans to launch a robotaxi service in China.Source: TechCrunch source, CnEVPost source
Ruigan Robotics obtains China's first dual gas-and-dust explosion-proof certification for a wheel-legged quadruped · industrial
The certificate, issued by the Shanghai Instrumentation & Automation Systems Inspection & Testing Institute, carries the rating Ex IIC T6 Gb & Ex IIIC T80℃ Db, covering both gas and dust explosive environments. IIC corresponds to the highest-risk gas category, including hydrogen and acetylene, with T6 requiring the equipment's maximum surface temperature not exceed 85°C; IIIC covers conductive dust, capping surface temperature at 80°C. Beyond the whole-machine certification, the company also obtained independent certification for its core explosion-proof components. The significance lies in the threshold this clears — in petrochemical, oil and gas, and coal mining settings, most robots to date have carried only single-gas explosion-proof or lower-tier certifications, locking them out of these environments. The product behind the certification is the Delta explosion-proof wheel-legged inspection robot, which supports 50kg of top-mounted payload, switches between wheeled and legged modes to cross ditches and steps, and carries onboard lidar, an infrared thermal camera, HD cameras, sonar, and gas detectors, paired with the Robo-AI petrochemical cloud platform for anomaly recognition.Source: iFeng Tech source
PaXini brings tactile sensing onto BYD's final assembly line · embodied ⚠️ Vendor claim
The two parties signed a deep strategic cooperation agreement bringing PaXini (Chinese tactile-sensing startup)'s multidimensional tactile sensing into manually operated key stations on BYD's final assembly line. BYD Group Vice President Luo Zhongliang and Board Secretary Li Qian attended alongside PaXini founder and CEO Xu Jincheng, with the agreement signed by Zhao Weibing, director of BYD's Future Lab, and Luo Xiaoheng, PaXini's Chief Strategy Officer. The press release explains why final assembly is the sticking point: stamping, welding, and painting handle model changeovers with flexible fixtures and vision-based positioning, switching over within two hours; but final assembly's hundreds of thousands of clip, adhesive, connector, and fitting operations require recalibration for every new model and still rely on veteran workers' feel for material hardness and deformation variance. At a takt time of one vehicle every 1.4 seconds, these operations have so far resisted being absorbed by vision-based systems. The release gives no numbers on units installed, stations covered, or any acceptance metrics — this is a cooperation signing, not a deployment milestone. Both parties say they will jointly advance the Ministry of Industry and Information Technology and State-owned Assets Supervision and Administration Commission's special program for real-world humanoid robot and embodied intelligence training.Source: China.com source
Realman's RealBOT stationed at Beijing Daoxiangcun to make mooncakes · embodied
Ahead of the Mid-Autumn Festival, Realman Robotics' wheeled humanoid robot RealBOT was stationed at the No. 0 treasure-hunt outlet of Beijing Daoxiangcun (well-known Beijing pastry brand), working alongside store staff to make Beijing-style mooncakes, with the in-store deployment running through the Mid-Autumn Festival. This is not their first collaboration — at WRC 2026, the two had already demonstrated the process of making Beijing-style "flaky pastry" five-kernel mooncakes using the GLN remote operation network. The explicit purpose of this round is to record pouring angle, mixing force, and unmolding timing as real-robot data to feed back into the model. Founded in 2018, the company self-develops all four of its core components in-house; its joint modules currently have a maximum annual production capacity of 500,000 units, with a target of reaching a million units annually by 2026.Source: Gasgoo source
Wandercraft signs 12 customers for Calvin-40, Renault plans to deploy 350 units in 18 months · humanoid ⚠️ Vendor claim
Calvin-40 is this French company's humanoid robot for material handling, with customers spanning automotive, logistics, retail, healthcare, and defense; the customer list has not been disclosed. The only publicly named customer is Renault Group, which plans to deploy 350 units within 18 months. Both the 12 customers and the 350-unit deployment figure come from the company's own press release and have not been independently confirmed.Source: GlobeNewswire source
Industry Developments
Draft revision of Road Traffic Safety Law adds dedicated autonomous driving chapter, dividing liability by "whether the function was activated" · autonomy
A draft revision of China's Road Traffic Safety Law has been submitted for first reading to the Standing Committee of the National People's Congress and is now open for public comment; it adds a new dedicated chapter, "Special Provisions for Autonomous Vehicles." This marks the first time China has clarified the legal status of autonomous vehicles, the conditions for operating them on public roads, and liability for violations at the national law level. Previously, relevant rules existed only in departmental regulations, normative documents, and national standards; while some localities had attempted their own legislation, its scope was limited and standards for allocating accident liability were inconsistent. The draft defines an autonomous vehicle as one that continuously performs the entire dynamic driving task in place of a human driver, within its designated operational design domain. Liability is allocated accordingly: for road traffic safety violations occurring while the autonomous driving function is activated, the manufacturer or importer bears responsibility for handling the matter; if the company believes the violation is unrelated to the autonomous driving function, the burden of proof falls on the company. Vehicles with the function deactivated, and vehicles with only assisted-driving functions, continue to be regulated as non-autonomous vehicles. The insurance provisions set only a framework — mandatory traffic insurance will apply to autonomous vehicles, with specific implementation rules to be set by the State Council. The draft also closes a marketing loophole, requiring companies to accurately disclose the vehicle's operational design domain and prohibiting product descriptions or commercial promotion that misrepresent it. Two scholars interviewed offered differing assessments. Zheng Fei, a professor at China University of Political Science and Law, argues that the legal nature of the "human-vehicle handover" remains undefined — neither takeover duration nor operating standards have been set, making liability for accidents during a handover still hard to determine, and civil compensation, product liability, and criminal liability will all need to be spelled out through supporting regulations. Zhang Li, deputy dean of the law school at the same university, takes a different view, arguing that the Road Traffic Safety Law is not currently the right vehicle for regulating civil liability at all — doing so properly would require adding manufacturers, operators, and system developers as liable parties alongside humans, and that "the timing may not yet be ripe."Source: Yicai source
UBTECH lands 1.5 billion yuan project in Shanxi, sending humanoid robots into coal mines · humanoid
On September 2, the Shanxi Comprehensive Reform Demonstration Zone signed an agreement with UBTECH and Beijing Hehao Ruisheng for the "Shanxi Embodied Intelligence Innovation Center and Industrial Energy Mining Special Robot Intelligent Manufacturing Project," with total investment of 1.5 billion yuan, to be built at Xiaohe New Industry Park. The project has two phases: the first is a data collection and scenario testing center, and the second is a UBTECH robot factory, aiming to become a core intelligent manufacturing base for robotics in North China. The choice of Shanxi comes down to coal. Shanxi's above-scale raw coal output was 1.305 billion tonnes in 2025, roughly 27% of the national total of 4.83 billion tonnes and the highest among all provinces. According to the Shanxi Provincial Energy Bureau in December 2025, the province had built 275 intelligent coal mines and 1,594 intelligent extraction faces, with intelligent mines accounting for 63% of production capacity, and underground substations and pump rooms now largely unmanned. What remains are inspection, maintenance, and auxiliary transport roles — high-risk jobs that still require mobility, exactly the segment special-purpose robots are meant to take on. Underground environments demand a level of explosion-proofing, dust resistance, reliability, and remote control beyond what a factory floor requires, making the case for paying for robots easier to argue than "humanoids on car assembly lines." Skepticism persists. According to an April 2025 report by Xinhua-affiliated Economic Information Daily, Boston Dynamics founder Marc Raibert has said that car manufacturing is already a strength of traditional industrial robots, and that humanoid robots can play only a limited role on the production line, more like a "mascot"; Abderrahmane Kheddar, a member of France's National Institute for Digital Science and Technology and professor at the University of Montpellier, likewise believes humanoid robots' advantages have yet to be fully demonstrated.Source: OFweek Smart Manufacturing (republished by Eastmoney) source
Inside a Beijing humanoid data facility: real-robot data runs 500–1,000 yuan per effective hour, and prices are still falling · adjacent
A reporter from STAR Market Daily visited the data and training facility of the Beijing Innovation Center for Humanoid Robotics. Spanning over 6,000 square meters in total, with roughly 5,000 square meters on the first two floors dedicated to data collection and training, the facility has built out more than thirty realistic scenarios across six major categories — home, retail, industrial, healthcare, and others — and has deployed over 150 sets of embodied robot equipment plus more than 150 sets of non-embodied capture equipment. The operating rhythm resembles a factory: collectors work 8-hour shifts, and after accounting for setting up scenes and props, get about 6 effective hours of recording; data is transmitted as it's collected, processed overnight by an automated pipeline, and delivered the next day — an internal cycle nicknamed "7+1." Annual capacity is 180,000 hours, and the facility has served over 100 clients. The most concrete figure in the piece is pricing: real-robot data costs roughly 500 to 1,000 yuan per effective hour and has already come down, while non-embodied data ranges from tens to one or two hundred yuan; the facility acknowledges it does not currently make money selling data. The person in charge pushed back directly on industry claims: "Some organizations say they'll reach ten million hours by year-end; others say they already have tens of millions of hours of non-embodied data — but how much of that is genuinely high-quality data?" Their standard is data that produces real improvement in a given scenario, likened to the truly valuable "corner cases" in autonomous driving. Under a three-stage quality check, the data delivered to one leading client in the first half of this year passed validation at over 95%. On the open-source side, the center's open-sourced RoboMIND dataset has recently surpassed 20 million downloads worldwide.Source: Cailianshe · STAR Market Daily source
XPeng's second-generation VLA 6.3.0 rolling out this month, model now retains 30 seconds of memory · autonomy
The new version begins rolling out to all Ultra and Ultra SE models starting in September, launching first on the G9L Ultra and Ultra SE, with Max models running a single Turing chip receiving a distilled version of the second-generation VLA, VLA Lite, the same month. Three modules correspond to three time horizons: Infini-VLA sets historical memory in the production version to 30 seconds, X-Foresight projects 6 seconds ahead, and Streaming Inference in between takes in input, runs inference, and outputs trajectories continuously, which XPeng says improves decision speed by 300%. Model parameter count has grown 3.5x, and data throughput per training run has reached 100 million video clips. The claim of "over 20x improvement in overall safety capability" comes from XPeng's own simulation and testing methodology.⚠️ Vendor claim The same architecture is also being ported to robotaxi use — internal testing has completed over 2,000 orders, and a vehicle equipped with the second-generation VLA recently obtained a remote-testing qualification for intelligent connected vehicles from Guangzhou authorities, allowing driverless-in-driver-seat testing on designated roads.Source: IFanr source
Yangtze River Delta G60 cities issue joint initiative on embodied intelligence quality standards · industrial
At the 9th Yangtze River Delta G60 Sci-Tech Innovation Corridor Quality Standards Conference in Hangzhou on September 4, participating cities released the "Yangtze River Delta G60 Cities Joint Initiative on Embodied Intelligence Industry Quality Standards," promoting standards across foundational/general technology, core technology, components, complete systems, and industry applications, and proposing cross-city collaborative governance of data. Hangzhou provided much of the basis for this initiative, having clustered over 700 robotics-related companies, with the local industry cluster's output value surpassing 106.8 billion yuan in 2025, and its market share in China for quadruped robots and humanoid robots exceeding 80% and 50% respectively. There has already been movement on standards: China's first group standard for maturity evaluation of quality management systems in the humanoid robot industry, T/ZRIA 002—2026, was initiated in November 2025 and took effect on June 19, 2026.Source: CNR source
CICC: China's industrial robot density has gradually surpassed that of the US, Japan, and Germany · industrial
CICC (China International Capital Corporation) Research released "The Impact of Factor-of-Production Shifts under New Quality Productive Forces on the Economy and Markets" on September 4, observing that China's industrial robot density has risen rapidly over the past decade or so and has in recent years gradually surpassed that of the United States, Japan, and Germany, though the report does not provide a specific density figure. Accompanying this trend, the synchrony between fixed-asset capital investment and new hiring has slowed at some manufacturing firms, with labor freed up from manufacturing flowing increasingly into services.Source: Guandian.cn source
Berg Insight: humanoid robot annual shipments to reach 26 million units by 2040 · humanoid ⚠️ Projected figures
This research firm projects a path from 16,000 units in 2026 to 26 million units in 2040, a 63.7% compound annual growth rate, corresponding to a market size growing from $890 million to $554 billion. In the same forecast, humanoid-robot-related cellular connections grow from 5,000 in 2025 to 50.6 million in 2040. The firm counts over 100 companies currently developing full-size humanoids, most still at the research, prototype, or pilot-deployment stage.Source: IoT Business News source
Hardware · Supply Chain
· Dasheng-1 reconfigurable dexterous hand: released September 4 by Heimai Technology (Chinese robotics component maker) jointly with LimX Dynamics, with 4 fingers and 11 active degrees of freedom, single-hand grip payload ≥24kg, repeat positioning accuracy ≤0.03mm, fingertip force ≥18N, and control frequency above 100Hz; the maker says it can switch between 7+ working configurations with zero downtime, replacing multiple gripper sets that would otherwise need to be swapped on a production line. source
· Black Sesame Technologies Aura module: Aubo Robotics (Chinese collaborative robot maker) has completed delivery of its first batch of Aura development kits in July, for use in collaborative robot development. Aura belongs to the SesameX platform, aimed at multi-legged robots and collaborative arms, offering 70 TOPS of heterogeneous compute; the platform spans three tiers from the 16 TOPS Kalos to the 700 TOPS Liora. source
· Uniview inspection vision modules: Q Technology (Group) says its subsidiary Uniview has partnered with DEEP Robotics and Weilan Robotics, with vision modules for inspection robots now shipping in volume. source
Top comments (0)