Today's highlights
· Unitree Robotics shares surged 460.34% on their first day on the STAR Market, closing with a market cap of 341.8 billion yuan
· Over 70 embodied AI training grounds built and operating nationwide in China; industry standard for dataset quality takes effect in November
· Shaanxi Automobile Heavy Trucks and Kargobot deliver the first global batch of 100 mass-produced L4 line-haul trucks
· Hexagon's AEON humanoid enters Schaeffler factory "gym," rollout planned for at least 1,000 units
· Dubai's robotaxi fleet passes 4 million kilometers, but the 144-vehicle fleet averages only about 50 rides a day
· Hydra-0 represents actions as pixel motion, cutting robot motion error by 90.4%
· Chery Robotics' subsidiary preparing IPO; about 110 police-use humanoids already deployed
1. Research progress
Hydra-0: representing "action" as pixel motion for cross-embodiment world models · world-model
One reason world models have struggled to generalize is that different robots' action spaces don't share a common representation. Hydra-0 proposes action flow — representing robot actions uniformly as pixel motion on screen, so data across embodiments, tasks, environments, and even different video-generation backbones can be fed into the same model. Under the best configuration, robot motion error drops 90.4% and object motion error drops 60.2% versus action-conditioned baselines; on RoboLab, the Pearson correlation between replayed success and real success reaches r=0.96 (making it usable for offline policy evaluation). The authors also found a "reverse mode": given a desired object motion transferred from human demonstrations, the model can infer compatible robot actions without expert robot data for that task.
Hongyu Li et al. · arXiv 2608.18077 source
QWM: world model only searches, doesn't train the policy, sidestepping compounding model bias · world-model
Past model-based RL trained policies or value functions directly on "imagined trajectories," and bias compounded as tasks got longer and scenes more complex. QWM repositions the world model — the policy and value function are trained only on real interactions, while the world model is used solely at inference time to run test-time search over imagined trajectories to select high-value actions. On Robomimic and LIBERO, both sample efficiency and final performance significantly exceed the previous strongest methods.
Perry Dong et al. (including Chelsea Finn, Dorsa Sadigh) · arXiv 2608.17163 source
FetchMan: humanoid mobile manipulation trained purely in simulation, 73.3% zero-shot success on real hardware · locomotion
Humanoids that walk while grasping face far higher real-world data collection costs than tabletop manipulation. The authors port the "sim training + sim2real" recipe common in locomotion to loco-manipulation, and identify a key step: cloning synthetic demonstrations alone hits a performance ceiling that more data can't break through — reinforcement learning with Flow-GRPO on a single sparse reward is required to break past it. The pipeline covers over 150,000 scenes, and the trained policy was deployed zero-shot to a real Unitree G1, achieving 73.3% success at approaching and grasping a single target object in unseen scenes.
Omar Rayyan et al. · arXiv 2608.17027 source
VLCP: no fine-tuning of the VLM — it rewrites its own control code within a single episode · manipulation
Turning frontier VLMs into robot policies usually requires fine-tuning them to output action representations not seen during pretraining, which erodes their reasoning ability. VLCP does the opposite: the VLM stays fully frozen and only writes a Python control function — no demonstrations, no fine-tuning. Every K steps it re-observes multi-view RGB and proprioceptive state and directly rewrites the failing code, rather than retrying the same policy or switching subtasks. Across 57 MuJoCo/RoboVerse tasks, this training-free policy achieves a combined success rate of 35.1%, versus just 3.5% for the same system when it only queries the VLM once per episode — a tenfold gap with non-overlapping confidence intervals across every task family, with the gains coming mainly from a 27.3% within-episode self-recovery rate after grasp failures.
Dhia Naouali et al. · arXiv 2608.16978 source
ManiGuard: task success doesn't mean safe success — 6–21% of "successful" trajectories violate safety specs · benchmark
Success rates for manipulation foundation models are climbing fast, but "whether success was achieved safely" is rarely evaluated rigorously. ManiGuard defines safety specifications independently of task success, using LTLf automata for runtime checking of physical predicates (rather than learned classifiers or LLM judges), running over 23,000 rollouts in simulation and on a real Franka arm: 6–21% of "successful" trajectories actually violate the safety specification. After fine-tuning with the accompanying 8,000 safety-annotated demonstrations, the safe-completion rate rose from near zero to 7.5–29.8%, but 21–42% of intervened trajectories still violate specs, and adding more data doesn't close the gap.
Yiyan Peng et al. · arXiv 2608.17386 source
Teach and Grow: treating deployment as continued learning to sidestep the "retraining tax" · vla
When an end-to-end VLA encounters objects, sensors, or embodiments outside its validation coverage, fixing a single error means re-collecting data, updating the policy, and rerunning regression tests — what the authors call a "retraining tax." TGL lets a multimodal agent crystallize a small number of successful demonstrations into reusable Skill Blocks (subgoal-oriented closed-loop behaviors), which it recombines in new scenarios, observes physical outcomes for, and reroutes when they deviate from intent; the skill library and experience memory accumulate continuously, and new tasks require no targeted policy retraining. It achieves SOTA on LIBERO benchmarks.
Chang Nie et al. (Shanghai Jiao Tong University) · arXiv 2608.17209 source
TAMP-Nav: the VLM only points at pixels, navigation success rate reaches 66.2% · autonomy
HuggingFace trending 37↑. Where existing approaches often force the VLM into an action space it never saw during pretraining, this paper instead has it "select a 2D pixel, then project it into 3D for a SLAM controller to handle," paired with a mechanism that triggers chain-of-thought only at key nodes and retains high-fidelity memory. Success rate on R2R-CE reaches 66.2%, with training using only 90,000 trajectories.
Hongyan Feng et al. · arXiv 2608.17512 source
Other papers today: LIBERO-VIFO finds that VLAs will follow "unauthorized" visual prompts in the scene even without a language instruction, exposing a new safety surface (arXiv 2608.17600 source); Prism-GRPO splits grouping by outcome, needing up to 56% fewer rollouts to reach target success rate (arXiv 2608.17423 source); UniReflex adds plug-and-play variable-impedance force control to frozen generative policies without fine-tuning the backbone (arXiv 2608.17432 source); EATR-Stereo designs proprioception-aware token routing for humanoid stereo headsets (arXiv 2608.17453 source); Scalix uses learned depth cues to recover true scale for monocular SLAM (arXiv 2608.17553 source); ControlledShifts re-splits trajectory prediction data into in-distribution/out-of-distribution partitions to standardize robustness evaluation (arXiv 2608.17882 source); V-JEPA4A replaces random masking with saliency-guided masking for self-supervised learning on driving video (arXiv 2608.17178 source); HODAgent is a System-2 agent for humanoids in service scenarios (arXiv 2608.17584 source).
Open source · tools · benchmarks
· PRISM industrial contact-rich dataset: over 25 categories of assembly-level manipulation (electronic component insertion/removal, conveyor sorting, etc.), 5,000+ trajectories totaling 45 hours of teleoperated demonstrations, with synchronized multi-view RGB-D, six-axis force/torque, tactile, and proprioceptive state; fills the gap left by existing datasets skewed toward "short-horizon, low-contact pick-and-place," now open-sourced (arXiv 2608.17962 source)
· Wuyi Vision (5911 Vision, Chinese startup) AperData embodied data platform: an integrated hardware-software solution combining the AperEgo capture rig and the AperOS data platform, launch priced at 5,100 yuan per set, moving quality checks to the capture stage to filter out invalid clips upfront; the vendor claims delivered data trajectories reach 99% physical consistency and production efficiency more than 10x that of real-robot teleoperation at equivalent cost source
· Xi'an Jiaotong University's Video2Knowledge: unveiled at RCIM 2026, claimed to teach robots new skills from a single smartphone-recorded video source
· Foxglove: adding AI tools to robotics data workflows, targeting log search and visualization source
2. Funding and deals
Unitree Robotics | STAR Market IPO first day | closed up 460.34% | market cap 341.8 billion yuan · humanoid
Following an August pre-listing bookbuilding round priced at a 219.23x P/E, Unitree Robotics formally listed on the STAR Market on August 19: it opened at 1,100 yuan/share (also the day's high, roughly 629% above the 150.80 yuan offer price) and closed at 845 yuan/share, up 460.34%, with turnover of 85.28% and trading volume of 23.16 billion yuan, giving a closing market cap of 341.8 billion yuan. The IPO aims to raise 4.202 billion yuan, of which about 3.572 billion yuan is earmarked for R&D, focused on robotics foundation models, hardware performance upgrades, and new product development. On financials, the company posted 2025 revenue of 1.699 billion yuan and net profit attributable to parent of 278 million yuan, with a 226.78% compound revenue growth rate over 2023–2025; in H1 2026 revenue was 1.152 billion yuan, up 48.54% year-on-year, and net profit attributable to parent was 274 million yuan (versus a loss of 32.02 million yuan in the same period last year), though non-recurring net profit fell 19.34% year-on-year to 244 million yuan, mainly due to a sharp rise in R&D and sales expenses. Worth noting is the demand structure: at this stage, its humanoid robot customers are still mostly universities, research institutes, and AI labs, purchasing through research budgets, while orders actually landing on industrial production lines remain limited; the company's proprietary general-purpose embodied foundation model is still in R&D testing and pilot deployment, and the prospectus explicitly states it has not been applied at scale to products during the reporting period. On the shareholder side, Meituan's pre-IPO stake was about 9.65%, diluted to about 8.68% post-IPO, giving it a paper gain of nearly 30 billion yuan at the closing market cap. The same day, China's A-share humanoid robotics sector broadly weakened, with Green Harmonic (harmonic drive maker) falling over 17%, and multiple stocks including Xingsen Technology and Fenglong (auto parts maker allied with UBTech) hitting the daily limit down.Source: STCN source
AiMOGA (Chery Robotics subsidiary) | preparing IPO | listing venue under discussion · humanoid
Zhang Guibing, head of Chery's AiMOGA and president of Chery International, told Reuters at the WRC venue that the company is preparing for a possible IPO and is in talks with several listing venues, without disclosing timing or location. The company, incubated in January 2025, says it has delivered over 3,000 robots globally (about 2,000 overseas) and operates in more than 60 countries and regions, aiming to raise annual deliveries to 10,000 units next year. Its police-use humanoid robot, launched this year, has already been deployed in multiple Chinese cities — about 110 units — for traffic direction, crowd guidance, and public safety outreach; Zhang said the hot, rainy Middle East and Southeast Asia are priority overseas markets. AiMOGA originally developed a bipedal humanoid for Chery dealerships and has since pivoted toward public service and companion-type products.Source: Reuters (via News.az) source
Motion Brain (眸深智能) | Pre-A+ round | nearly 500 million yuan · world-model
Following a nearly 100 million yuan top-up to its Pre-A round in July, Motion Brain has closed a Pre-A+ round of nearly 500 million yuan, jointly invested by the Shenbao Yiben Fund, Orient Securities, the Shaanxi High-tech Industry Investment Fund, along with Anyu Capital, Tianmeng Investment, and Jianyuan Tianhua, with existing shareholders Chuanghehui, Xuhui Capital, and Gengxin Capital adding oversized follow-on stakes; the company says its valuation grew more than 10x in the first half of the year. The team was founded by Fudan University professor Chen Tao, former Intel China chief scientist Zhang Yimin, and serial entrepreneur Mu Zelin, and pursues a "world action model" approach — its earlier MotionGPT decomposed actions into roughly 3,000 "action primitives," and this year it launched the spatiotemporally unified STI-WM. Commercially, the company says audited revenue collected reached the low tens of millions of yuan in 2025 and has already hit 30 million yuan in H1 2026, having delivered robot products to a listed property management firm and an A-share-listed sanitation industry leader.Source: PEdaily source
Chenxing Technology (formerly Atom Robotics) | filed for Hong Kong listing | Huatai International sole sponsor · industrial
Parallel robot maker Atom Robotics, after rebranding as "Chenxing Technology," filed a prospectus with the Hong Kong Stock Exchange under the 18C specialist technology company route, with Huatai International as sole sponsor; the company is described as China's largest parallel robot brand, with shareholders including Yang Haorong, Shenzhen Capital Group, and Lenovo.Source: Sohu Finance source
Fenglong (auto parts maker, controlled by UBTech) | H1 results swing to loss | net loss of 2.0151 million yuan · hardware
After UBTech took a 29.99% stake at 17.72 yuan/share in March this year (later raised to 43.01%), Fenglong reported its first half-year results under the new ownership: revenue of 267 million yuan, up 6.48% year-on-year, while net profit attributable to parent swung from a profit of 15.3141 million yuan in the same period last year to a loss of 2.0151 million yuan, which the company attributed to foreign exchange losses, higher aluminum ingot prices dragging down gross margin (overall gross margin fell 4.16 percentage points to 18.35%), and depreciation from a subsidiary's factory being capitalized. On the robot components business that the market is watching most closely, the operating entity Xinfenglong Technology, established in Shenzhen in June, has not yet been capitalized and has not begun operations, contributing no revenue by period end; the company expects related-party transactions with UBTech in 2026 not to exceed 82.5 million yuan, and as of the disclosure date these had also not yet generated revenue. The stock hit the daily limit down on August 19, closing at 49.33 yuan.Source: International Finance News source
Luming Robotics | Series A+ | Fudan University venture fund · embodied
Luming Robotics closed a Series A+ round funded by the Fudan University venture capital arm, amount undisclosed; the company showcased NexCore-related products at WRC the same day.Source: DoNews source
3. Commercialization and deployment
Shaanxi Automobile Heavy Trucks and Kargobot deliver first global batch of 100 mass-produced L4 transport robot platforms · autonomy
On August 18 in Ordos, Inner Mongolia, Shaanxi Automobile Heavy Trucks and Kargobot completed delivery of the world's first batch of 100 mass-produced transport robot platforms and signed a full-lifecycle strategic cooperation agreement. Unlike the common pattern of "technical cooperation plus small pilot batches," the two companies have integrated vehicle manufacturing, autonomous driving technology, operations, and after-sales support into a unified "vehicle + autonomous driving + service" model. Line-haul logistics is one of the few L4 commercialization scenarios where the economics clearly pencil out — high labor costs, intense workloads, and concentrated safety risk — and a hundred-unit-scale delivery is a concrete milestone in China's shift from R&D validation to scaled operation for commercial-vehicle autonomous driving.Source: Sohu (reprint) source
Hexagon's AEON humanoid enters Schaeffler's "humanoid gym," rollout planned for at least 1,000 units · humanoid ⚠️ planned figures
Hexagon Robotics and Schaeffler announced that the AEON humanoid robot has entered Schaeffler's Humanoid Gym in Germany, the next step following the strategic partnership the two companies established in April, pointing toward a deployment plan of at least 1,000 units of AEON over the coming years. The approach of this dedicated industrial "train–validate–deploy" environment is notable: AEON refines and validates policies for representative manufacturing tasks drawn from Schaeffler's factories through imitation learning and repeated execution, both minimizing production-line disruption and simultaneously building Schaeffler's internal capability to operate, train, and integrate humanoids. Schaeffler's dual role is another variable — it is both the end user and a supplier of actuator technology to AEON. The two companies say they will extend AEON to more use cases in Schaeffler's manufacturing processes over the next six months.Source: Robotics & Automation News source
Dubai's robotaxi fleet passes 4 million kilometers, but the 144-vehicle fleet averages only about 50 rides a day · autonomy
Dubai's Roads and Transport Authority (RTA) disclosed that its robotaxi service has driven over 4 million kilometers since launch, completing 7,613 passenger trips with a 97% customer satisfaction rate. But the denominator tells a more interesting story: the fleet totals 144 vehicles, of which 48 are in active service, averaging about 50 rides a day — most of the mileage comes from safety-driver-supervised validation operations rather than scaled passenger service. RTA officials said this provides a basis for evaluating Dubai's goal of having 25% of trips completed autonomously by 2030, with the service transitioning from safety-driver trials toward fully driverless operation. Dubai's three partners are Baidu's Apollo Go, WeRide, and Pony.ai.Source: Dubai Government Media Office source
Brightpick adds 44 more picking robots for Europe's largest pharmacy chain, bringing the total to 74 · industrial
Brightpick announced an expanded partnership with Dr. Max, Europe's largest pharmacy chain group, deploying 44 additional Autopicker autonomous mobile picking robots across the Czech Republic and Slovakia, bringing the group's total operating robot count to 74 once complete. This is a fully proven scaling curve: the partnership began with 9 units at a Prague warehouse in late 2023, gradually expanded to 30 units, and that warehouse now processes up to 20,000 items a day with almost no manual intervention — manual picking labor down 95%, storage capacity up 40%, and picking error rate near zero. The two new sites in Ostrava and Bratislava will each deploy 22 units, handling pharmacy/hospital restocking and online order picking respectively, with completion due in the second half of the year.Source: Sohu (reprint) source
South Korea's Robros deploys nine bipedal humanoids into semiconductor cleanrooms and cold-chain warehouses, all under paid contracts · humanoid
Through the Korea Robotics Industry Association's "Humanoid Robot Field Verification Support Project," Robros is supplying 9 units of its proprietary bipedal humanoid IGRIS-C to consortia across semiconductors, automotive manufacturing, and fresh-food cold-chain logistics — 3 of the 10 consortia selected for the project chose this platform, and the supply is entirely under paid contracts rather than trials. Including prior sales, IGRIS-C has sold 18 units cumulatively and produced 36 units total, with the remaining 18 used for R&D, technology demonstrations, and customer testing. The robot stands 154 cm tall and weighs 56 kg, with end effectors swapped by scenario: a five-fingered hand for general-purpose tasks and an electromagnetic gripper for handling metal parts.Source: Venture Square source
Pony.ai to add hundreds more vehicles overseas this year, targeting 500–1,000 heavy trucks within 2–3 years · autonomy
Following its Q2 robotaxi revenue growth of 691% year-on-year (previously reported), Pony.ai CEO James Peng added detail on international expansion: the company will deploy "hundreds more" vehicles outside mainland China by year-end, with overseas deployment reserves reaching 4,000 vehicles, and a global fleet target of 3,500 vehicles by the end of 2026 (up from 1,975 as of June 30). He said openness to autonomous driving in Europe has clearly increased over the past year, with pilot zones gradually opening up. On heavy trucks, the company plans to deploy 500–1,000 fourth-generation autonomous trucks over the next 2–3 years for line-haul freight, bulk cargo transport, and port logistics; the key to the expansion decision is cost — the fourth-generation truck's autonomous driving hardware costs about 70% less than the previous generation, with the vehicle designed for a service life of 20,000 operating hours or 1 million kilometers. It had 200 trucks on the road as of the end of 2025.Source: South China Morning Post source; CleanTechnica citing CnEVPost source
4. Industry developments
CAICT: over 70 embodied AI training grounds built and operating nationwide in China, dataset quality industry standard takes effect November 1 · adjacent
The China Academy of Information and Communications Technology (CAICT) released its "Embodied AI Training Ground Research Report (2026)" on August 18: China has built and put into operation over 70 embodied AI training grounds nationwide, with another 40-plus under construction or planned, and the supply chain has taken initial shape. Geographically, the layout relies on three core clusters — the Yangtze River Delta, Beijing-Tianjin-Hebei, and the Pearl River Delta — already covering more than half of China's provincial-level regions, and extending into third-tier and lower cities, diverging from the pattern of traditional AI industry clustering heavily in first-tier cities. A stricter development is the standard: the industry standard "Embodied AI Dataset Quality Requirements and Evaluation Methods," drafted by CAICT together with over 40 organizations, takes effect November 1, setting rules across data production operating norms, organizational mechanism safeguards, and quality evaluation indicators, with the evaluation indicators covering eight dimensions including completeness, consistency, diversity, and authenticity. CAICT says this marks a shift in embodied AI dataset construction from "scale-oriented" to "quality-oriented." The report also cites figures from a Zhongguancun-based (Beijing tech hub) company: its training ground has delivered 1.5 million hours of human video data, while training a high-quality embodied foundation model often requires tens of millions or even hundreds of millions of hours. A joint notice from the Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission, the "2026 Special Action Notice on Real-Scene Field Training for Humanoid Robots and Embodied AI," is focused on building a closed loop of "real-scene field training – data accumulation – product iteration – scaled deployment" across the industrial, service, and specialized robotics sectors.Source: World Internet Conference source
JD.com unveils robotics strategy: 10 billion yuan in resources by 2028, plans to collect over 10 million hours of real-world scenario data in two years · adjacent ⚠️ planned figures
As a global strategic partner of WRC 2026, JD.com unveiled a robotics strategy spanning supply chain, service, and technology. On supply chain, it will invest 10 billion yuan in resources in robotics by 2028, aiming to help 100 brands each achieve independent sales exceeding 1 billion yuan; it currently has self-operated partnerships with over 200 robotics brands and core component makers, and its European platform Joybuy has onboarded over 50 Chinese robotics brands. On service, it plans to build after-sales support covering over 100 countries within five years, along with 80 RoboBase robot service centers. On technology, JD Cloud plans to collect over 10 million hours of real-world scenario data within two years, and its embodied AI platform JoyInside targets over 10 million connected terminal devices by year-end. JD.com also launched a standardized robot battery product, unifying dimensions, interfaces, communication protocols, and safety standards, already partnering with over 20 brands — at a time when components remain highly non-standardized, this kind of interface unification move may have industrial impact no smaller than a full robot launch.Source: Chaoqi Net source
TrendForce: China's humanoid robot market to reach about 15 billion yuan in 2026, WRC holds first "Procurement Day" · humanoid ⚠️ estimated figures
TrendForce estimates China's humanoid robot market will reach 15 billion yuan in 2026, and will maintain at least 60%-plus growth in 2027 as mass production scales and use cases expand. It flags one signal: this year's WRC held its first-ever "Procurement Day" for concentrated procurement negotiations and supply-chain matchmaking, indicating the industry's center of gravity is shifting from technology demonstration toward actual demand. The firm also compiled figures from several manufacturers — Agibot (Zhiyuan Robotics, Chinese humanoid startup) had cumulatively produced 15,000 units as of June 2026; UBTech's products have been adopted by customers including Airbus, with U1 pre-orders reaching 13,361 units by the end of June; Galbot (银河通用, Chinese humanoid startup) partnered with CATL and won a 236 million yuan embodied AI equipment procurement bid. Its judgment: beyond the show floor, what actually matters is which products have confirmed customers and are beginning batch delivery, and whether initial customers convert into repeat orders.Source: TrendForce (via IT之家/Sina Finance) source
Momenta partners with XHEART and QNX to build a production-grade autonomous driving platform, passes ASIL D certification · autonomy
Momenta (6880.HK), XHEART (新芯航途, Chinese automotive chip startup), and BlackBerry's QNX announced a joint production-oriented autonomous driving platform: Momenta contributes the full-stack algorithms, XHEART contributes the automotive-grade SoC X7, and QNX contributes a safety operating system based on SDP 8.0. The joint solution has passed TÜV Rheinland's ISO 26262 ASIL D highest-level functional safety certification, providing a compliance foundation for automakers developing autonomous-driving vehicles for international markets, particularly meeting access requirements for regulated markets such as Europe's UN R171 (Driver Control Assistance Systems, DCAS). As regulations on advanced driver-assistance systems tighten across major markets, deep integration of base software, chips, and algorithms is shifting from an engineering problem to a prerequisite for going overseas.Source: LeiPhone source
IFR: US industrial robot installations reached 38,000 units in 2025, up 11%, with food industry surging 30% · industrial
International Federation of Robotics (IFR) data shows US industrial robot installations reached 38,000 units in 2025, up 11% year-on-year, with the recovery driven mainly by non-automotive sectors such as food — food industry installations surged 30% to about 3,000 units, now on par with metalworking, machinery, and electrical/electronics; automotive remains the largest application at 13,500 units, down slightly 1% year-on-year. In robot density, US manufacturing has 307 robots per 10,000 employees, ranking 8th globally (up two places from last year), behind South Korea's 1,220, Germany's 449, and Japan's 446, but ahead of China's 166. Scale tells a different story, though: China installed 295,000 units annually in 2024, accounting for 54% of the global total, and IFR estimates its 2025 installation volume at about ten times that of the US. The Association for Advancing Automation (A3) has formally submitted a "National Robotics Strategic Vision" to lawmakers, calling for a federal robotics office and a national commission.Source: Sina Tech source
Germany's VDMA calls for Europe to build up local supply chains in humanoid robotics · humanoid
On the very day Unitree's shares surged on its IPO, the German Mechanical Engineering Industry Association (VDMA) publicly called for Europe to move faster and build a local supply chain in humanoid robotics.Source: Reuters source
Guangxiang Technology's Phi-Bot X1 demonstrates a single unit autonomously cycling between heterogeneous workstations · industrial ⚠️ vendor claim
Guangxiang Technology unveiled its industrial-grade embodied AI robot Phi-Bot X1 at WRC, with a live demonstration of the same unit autonomously switching between two different types of workstations — welding material handling and mobile quality inspection — and cycling through work without human intervention, which the company calls a first for the industry. Traditional industrial robots are fixed to a single workstation and require manual intervention to switch lines, so this approach directly addresses the pain point of flexible, small-batch, multi-variety production. One caveat worth noting: this was a live demonstration on the show floor rather than routine production-line operation, and the company has not disclosed a mass-production timeline — this gap between "can demonstrate" and "can deliver" is exactly what buyers should be probing at this year's WRC Procurement Day.Source: Guandian.cn source
Physen AI partners with Yuhuatian to build a dataset of tens of millions of hours of manual-labor interaction data · adjacent ⚠️ planned figures
Physical interaction data infrastructure company Physen AI (真觉万象) has reached a strategic partnership with Yuhuatian (玉禾田, 300815.SZ), a leading Chinese urban services operator, and its subsidiary Xiaorun Zhike, to jointly build a dataset of tens of millions of hours of embodied interaction data drawn from real manual labor, leveraging Yuhuatian's operational network covering over 200 cities and nearly 100,000 frontline workers. The entry point is well-chosen: scenarios like street sweeping, waste sorting, and public restroom cleaning are highly repetitive, subject to strong disturbances, and represent exactly the kind of task where robot substitution has the highest commercial value while being hardest to replicate through simulation — the changing grip force on a broom or the body's balance adjustments on wet, slippery ground can't be simulated. On hardware, the setup uses Physen's EGO "three-ring" solution (a head ring with binocular vision positioning plus a dual-arm ring with six-axis force sensing), with total hardware cost held to the low hundreds of dollars, allowing data to be captured automatically as workers do their jobs with almost no disruption to existing workflows; it also replaces traditional EMG muscle-electrical signals with FMG force-myography, filling a data-collection blind spot for static grip and passive touch.Source: Stock Star source
Hardware · supply chain
· Zhaowei Machinery & Electronics: H1 2026 revenue of 816 million yuan, up 3.8%; net profit attributable to parent of 74.6457 million yuan, down 34.1%, non-recurring net profit down 44%, with a net margin of about 6.6% — R&D headcount expansion and production-line construction costs for the dexterous-hand and other robotics businesses have already been booked while revenue has yet to scale, and the financial report doesn't separately disclose dexterous-hand revenue; this comes against a current P/E of 146x and market cap of 21.9 billion yuan source
· Black Sesame Technologies × Midea Meichuangxi: signed a framework agreement to supply Chinese-made chips for robotics and appliances to Midea's chip application platform, entering via industrial and humanoid robots; its SesameX embodied AI computing platform includes the Kalos/Aura/Liora tiers, with compute ranging from 48 TOPS to nearly 600 TOPS source
· Lingxin Qiaoshou (dexterous-hand maker) × Orbbec: upgraded to a strategic partnership around "hand-eye coordination," with Lingxin Qiaoshou contributing scaled mass-production and real-world dexterous manipulation data, and Orbbec contributing its fully self-developed 3D vision modules; Orbbec says it has served over 5,000 enterprise clients, including 1,600 robotics customers source
· Basic Semiconductor (09971.HK) × UBTech (09880.HK): the two companies reached a strategic partnership source
Top comments (0)