Today's Highlights
· China's Supreme People's Court issues its first judicial guidance on AI-related cases with 24 provisions, automakers and dealers held liable for damages caused by false advertising
· Agility's SPAC filing reveals the numbers: 2025 revenue of $1.8 million, operating loss of $140 million
· Among 7 listed Chinese robotics companies' half-year reports, only Unitree Robotics (Chinese robotics maker) posted positive net profit attributable to shareholders
· Orbbec (Chinese 3D vision company) moves toward a Hong Kong listing as Q2 net profit falls to just RMB 10.64 million, down 70% year-on-year
· Meta trials a dual-arm robot for swapping server cables at its Iowa data center
· South Korea's tech and SME ministries team up to start with AMR/AGV adoption consulting for small and mid-sized factories
· SENAD (Chinese industrial robotics startup) secures nearly RMB 200 million in new funding, its fourth round in six months
Paper Roundup
Turning tactile prediction into in-execution correction lifts real-robot success rate from 22% to 64% · manipulation
Visual prediction can't see contact. Treating future tactile images as an extra view to predict recovers only a third of the achievable gain, the authors find, because of a timing mismatch: the prediction happens before execution, but tactile feedback only arrives during execution. TacPAC has the base model plan an action chunk, then caches the predicted contact that the plan depends on; a tactile expert compares each new frame's observation against the cache and corrects the not-yet-executed action, with each correction pass costing 20.7 times less than regenerating the action chunk. Across five real-robot tasks covering precision insertion, fragile object handling, object reorientation, and long-horizon manipulation, average success rate rose from 22% for the vision-only base model to 64%; code is open-sourced.
Zipei Ma et al. · arXiv 2609.05266 source
Rephrasing alone can flip a reward from failure to success on the identical trajectory · benchmark
VLMs are increasingly used as reward functions for robot learning, a role that demands rewrite invariance: semantically equivalent goal descriptions should score the same trajectory identically. Using 2,390 real-robot trajectories with ground-truth progress labels, plus 21,673 verified lexical/syntactic/goal rewrites, the authors built ROBORMBENCH and found that rewrite instability is common and severe across both closed- and open-source models, worsening as rewrites diverge more — scaling parameter count or adding explicit reasoning doesn't fix it. Only reward models specifically trained with trajectory supervision were notably more stable.
Wonje Jeung et al. · arXiv 2609.05401 source
LIBERO success rates are approaching 100%, but no one has tested recovery from a fall · benchmark
Missed grasps, collisions, objects knocked askew — unavoidable in real interaction, yet existing benchmarks sidestep them by starting from preset initial states. The authors collected real execution failures from SOTA embodied models and built over 1,000 scenarios across four tiers — action retry, action adjustment, object-state recovery, and environment recovery — evaluating spatial understanding, object-structure reasoning, interaction comprehension, and topological reasoning.
Lin Liu et al. · arXiv 2609.05178 source
MINT: world-frame bimanual trajectories directly from egocentric video, with a 1,021-hour dataset · perception
Recovering camera and hand motion in world coordinates from egocentric video previously required a multi-stage pipeline — camera motion, depth, hand reconstruction, trajectory optimization — that's costly and can't jointly model camera and hand. MINT uses a shared spatiotemporal video representation to jointly predict camera trajectory, camera-frame hand state, and per-frame hand presence, then applies an explicit coordinate transform to output world-frame hand motion, transferring zero-shot to unseen egocentric datasets. The accompanying open-source annotation pipeline, EGOPIPELINE, converts public egocentric video into structured camera and hand trajectory supervision; the team also released a 1,021-hour egocentric trajectory dataset along with model, training, and inference code.
Zijie Zhu et al. · arXiv 2609.04958 source
Real-robot online RL enters VLA post-training, averaging 98.3% success across nine chemistry tasks · vla
Pretrained VLAs generalize broadly but become unreliable on tasks demanding precision and repeatability. VLA-Precision uses asymmetric co-bootstrapping across time scales for post-training: early on, human-intervention-guided behavior learning quickly ramps up performance; once autonomous experience accumulates, global return propagation and local preference ranking progressively calibrate value estimates to suppress policy drift. The accompanying ACoB-Stream architecture handles invariant-state decoupling and on-demand streaming, boosting throughput and compute efficiency by up to 10.9 times. Across four robot embodiments and nine high-precision chemistry tasks, average success rate reached 98.3%, at 45.8 minutes per task.
Chenyu Su et al. · arXiv 2609.04355 source
Moving the cost of causal reasoning to training time, zero extra overhead at deployment · vla
Adding causal reasoning to a VLA does improve manipulation, but the cost lands at inference time: generating reasoning tokens or rolling out future states at every step, which snowballs on long-horizon tasks. Latent Semantic Scaffolding is an auxiliary loss that only operates during human-demonstration pretraining, using a small projection head to align the VLA's action-token representations with text embeddings of physical reasoning rationales; the projection head is discarded at inference, leaving the base policy unchanged. The key finding concerns alignment granularity: aligning each action token to the rationale for its own manipulation phase transfers far better than aligning to a pooled embedding of the whole episode — representation probes show the former roughly doubles the backbone's phase-wise separability, while pooled alignment overfits to the training tasks.
Andrew Ting Yan Li et al. · arXiv 2609.04893 source
TourPhysics: teaching world models to distinguish "looking" from "acting" · world-model
Moving the camera only reveals new surfaces, whereas intervention changes object motion, contact, and deformation — but most existing video world models are driven by appearance priors and tend to lose physical and spatial consistency over long horizons. TourPhysics starts from a single image and a declarative physics configuration, separating deterministic simulation from video generation: simulator state, geometric evidence, generator control, and appearance memory are each handled independently. For each action, the simulator first computes a bounded physical and camera trajectory, then generates the corresponding observation. On simulator-defined camera tours and object manipulation, it tracks the prescribed trajectory more closely than comparison methods and shows less appearance drift on long-horizon revisits.
Xin Zhang et al. · arXiv 2609.04911 source
Strict oscillatory modes extended from an isolated curve to fill the whole configuration space · locomotion
Strict nonlinear normal modes give conservative mechanical systems a well-behaved family of oscillations, but the engineering catch is that this usually holds only along a single isolated curve in configuration space. The authors construct a potential function compatible with the geodesic flow so that strict modes hold across the entire space — a result bearing on how elastically actuated and legged systems can oscillate more efficiently. The paper has been cited 5 times, the only paper in today's batch with a citation signal.
Arne Sachtler, Alin Albu-Schäffer · arXiv 2609.04817 source
Other papers today: RoboSPA expands 56 base tasks across 10 categories into 280 difficulty variants using 527K multi-embodiment trajectories, specifically testing VLAs' fine-grained spatial reasoning and long-horizon process planning (arXiv 2609.05324 source); FailureSpot generates weak supervision from unlabeled action chunks for timestamp-level VLA failure detection (arXiv 2609.04277 source); CoLMIN has multiple vehicles negotiate through multi-decision LLM paths to reach stable consensus (arXiv 2609.04807 source); a first attempt at estimating multi-person 3D pose from acoustic signals alone (arXiv 2609.04902 source); CrossDepth uses geometry-constrained attention for generalizable multi-view surround depth estimation, outperforming self-supervised SOTA on cross-domain evaluation across DDAD and nuScenes (arXiv 2609.05397 source); NavArena automatically converts fixed 3DGS reconstructions into goal-directed visual navigation benchmarks (arXiv 2609.04602 source); AquaBEV uses 3D imaging sonar as geometric supervision to predict BEV occupancy from a single underwater RGB image (arXiv 2609.04411 source); One Word, Different Action examines both invariance and sensitivity of real-robot decisions using task-preserving and task-changing instruction pairs (arXiv 2609.05260 source); Dressing in Motion has an assistive dressing policy keep pace with arm movement (arXiv 2609.04759 source); APEX-RBD explores a mixed-precision design space for rigid-body dynamics accelerators (arXiv 2609.05161 source).
Funding & Deals
Agility Robotics ｜ Pre-SPAC-merger financial disclosure ｜ $2.5 billion valuation · humanoid
2025 net sales were $1.8 million, with an operating loss of $140 million; operating expenses rose from $71 million in 2024 to $111 million. Following its August merger progress with Churchill Capital Corp XI, Agility's S-4 filing laid a humanoid robot company's books bare for the first time: Digit has been deployed at 9 customer sites, logging over 65,000 cumulative operating hours, with signed multi-year Digit v5 contracts exceeding $300 million in order value. The deal values Agility at $2.5 billion and is expected to raise over $620 million in total, roughly $420 million from the Churchill trust and $200 million from a PIPE led by Foxconn. Against 2025 revenue, that values the company at roughly 1,400 times annual sales. This is the first time a humanoid robot manufacturer will have to face public markets with quarterly reports — until now, all figures on delivery pace have come from the manufacturer's own statements.Source: The Robot Report source
SENAD (Chinese industrial robotics startup) ｜ new round ｜ nearly RMB 200 million · industrial
The round was funded by Wuliangye Fund, ZMVC (Sino-US venture capital firm), Xingzheng Investment, and Zhechuang Technology — SENAD's fourth round in six months. The company builds autonomous loading and unloading systems for the "last 20 meters" of industrial logistics; its flagship product, iLoabot-M, recently updated to version 2.0, expanding from single-scenario to multi-SKU complex scenarios. Earlier this year it released a loading/unloading vertical physics engine, Senad Robot Insight-World V3.0, used to predict physical outcomes before a grasp. Deployments concentrate in alcoholic beverages, tobacco, fast-moving consumer goods, pharmaceuticals, and high-end manufacturing, with overseas channels covering Europe, the US, and Japan. ZMVC founding partner Huang Zhiyi said scaled commercial deployment is becoming the yardstick for measuring embodied AI companies' real value.Source: ChinaVenture source
Bear Robotics ｜ Pre-IPO ｜ KRW 300–400 billion ｜ targeting roughly KRW 2 trillion valuation · adjacent ⚠️ Unverified/rumored
Citing unnamed investment bank sources, the Korea Economic Daily reports that Bear Robotics, the US-based delivery robot company controlled by LG Electronics, is preparing for a Nasdaq listing, having selected Bank of America to arrange financing and approached domestic and international private equity firms to participate, with proceeds earmarked for robotics R&D. The pre-IPO round is sized at KRW 300–400 billion, targeting a valuation of roughly KRW 2 trillion. South Korean robotics stocks strengthened this week after the government's 2027 budget proposal sharply raised robotics spending, extending support from technology development to procurement and mass production.Source: Korea Economic Daily source, Sina Finance source
Orbbec (Chinese 3D vision company) files for H-share listing in Hong Kong · hardware
Just after shedding the "U" designation on Shanghai's STAR Market, the company is opening a new fundraising window. 2025 was the first profitable year since listing for this 3D vision supplier, with revenue of RMB 941 million, up 66.66% year-on-year, and net profit attributable to shareholders of RMB 128 million. The picture shifted in H1 2026: revenue of RMB 438 million, up just 0.49% year-on-year, and net profit attributable to shareholders of RMB 41.64 million, down 30.82% year-on-year; Q2 revenue was RMB 235 million, down 3.97% year-on-year, with net profit of RMB 10.64 million, down 70%. BOCOM International accordingly cut its three-year revenue forecast by 19% to 24%. The company holds over 70% market share in 3D vision for service robots in China, and its sensors are used by the top four humanoid robot makers by global shipment volume — Zhiyuan Robotics (Chinese humanoid startup), Unitree, UBTech, and Leju Robotics (Chinese humanoid startup). According to Interact Analysis data, it holds a 46% share of 3D vision for industrial mobile robots in South Korea, versus 32% for RealSense. The timeline of capital moves is hard to parse: a month before the Hong Kong filing, the company had just completed a roughly RMB 1 billion A-share private placement; on the same day it announced the filing, it also disclosed plans to use RMB 900 million of idle proceeds to buy wealth management products, while controlling shareholder Huang Yuanhao had sold over RMB 300 million of shares on the secondary market in cumulative reductions in June.Source: Investor's China source
Zichuang Zhixie (Shanghai) ｜ new round ｜ tens of millions of RMB · industrial
Yida Capital participated in the round. Founded in 2025, the company builds models specialized for construction machinery operation; its self-developed iDM integrated model packages a VLM, a VLA, and a WAM world model into a "big brain + small brain" dual-system architecture, targeting loaders and excavators rather than mining trucks. Founder Liu Haiquan's view is that conventional autonomous driving emphasizes passive perception and adapting to a fixed environment to complete movement, whereas earthmoving actively reshapes the physical environment — the shape of a material pile keeps changing, making environmental understanding and dynamic decision-making more complex than point-to-point mining truck transport. By comparison, Yikong Zhijia (Chinese mining autonomy company) had deployed 2,580 active driverless mining trucks by the end of 2025, holding 55.5% of the market for driverless mining solutions in China. Zichuang Zhixie has signed strategic cooperation agreements with XCMG Hanyun and Shantui.Source: Sina Technology source
Lindong Technology (Beijing) ｜ Pre-A round ｜ over RMB 10 million · adjacent
Jinqiao Fund led the round, with existing investor Miracle Plus Ventures (Chinese startup accelerator) participating and Shineng Capital serving as exclusive financial advisor. Founder Chu Yichen is a doctoral student at Northeastern University; the team builds underwater bio-inspired robots using an underactuated structure to achieve "one drive, multiple motions," cutting the cost of core drive components by roughly 60%. The company is launching first in the education market: after finalizing its mass-production process in April 2026, it shipped over 1,000 education robot units within three months, has developed 23 underwater robot products to date, and plans to launch its first consumer-grade underwater tracking-shot robot in the second half of the year.Source: ChinaVenture source
Commercialization & Deployment
Meta puts a dual-arm robot into a live server room to swap cables · adjacent
According to WIRED, Meta's data center automation has moved past warehouse logistics and material handling into live server environments, testing maintenance actions like server restarts, cable swaps, and component reinsertion. The clearest test site so far is Altoona, Iowa, where two Watney Robotics dual-arm systems have been dedicated to cabling since June 2025, alongside tests of a Kinova Gen3 arm for power-cycling and a human-triggered remote button device. Cabling work looks simple but actually requires identifying the correct port in a dense space, positioning both arms, and applying just the right amount of force without disturbing neighboring cables. One worker told WIRED that if a cable-swapping robot really succeeds, it could take over about 80% of a technician's workload; TechRadar noted this figure reflects that worker's concern about automation, not a target set by Meta.Source: Global Sources, citing WIRED source
Hyundai WIA builds South Korea's first fully autonomous forklift, headed for Kia's Hwaseong plant in May 2027 · industrial
Hyundai WIA says this is the first time a South Korean company has built a forklift capable of autonomous loading and unloading. The vehicle navigates the plant autonomously using lidar, vision sensors, and safety scanners, calculating optimal routes via the factory's digital twin; it carries loads of up to 4 tons at a top speed of 6.5 km/h, with safety scanners on all four sides stopping the vehicle immediately upon detecting an obstacle. The company says the hardest part is loading and unloading: vision sensors must pinpoint each pallet's exact position, accounting for offsets from truck-bed sag under load and uneven ground, and the vehicle must adjust its posture for loading into containers on ramps. Deployment is set for May 2027 at Kia's AutoLand Hwaseong assembly plant in Gyeonggi Province. The vehicle will join Hyundai WIA's existing AMR and parking robot lineup under its H-Motion brand.Source: The Korea Times source
Zhiping Fangyuan (Chinese robotics startup) robot baristas make hundreds of cups a day in Beijing, Shanghai, and Shenzhen · embodied ⚠️ Manufacturer's own claim
Grab a cup, recognize the button, operate the machine, deliver the drink — one cup per minute. The company's general-purpose robot, AlphaBot 2, is fitted with 6 cameras covering the chin, chest, rear, and base, plus 1 lidar for 360-degree perception; the company says the entire coffee-making process is driven by its self-developed embodied foundation model, pre-trained on common commercial coffee machine models. PR lead Ge Zhenwei said this kind of "robot barista" deployment is already running routinely in Beijing, Shanghai, and Shenzhen, steadily producing hundreds of cups a day; on the day of the interview, the robot made nearly 100 cups that morning with zero errors. The company recently also placed its Aibao robot as a bartender in a bar in Hong Kong's Lan Kwai Fong district. The claimed uptime and cup counts describe a capability demonstration, not per-store unit economics, failure rates, or how often human intervention is needed — those figures remain known only to the company.Source: Shenzhen Commercial Daily · Readhub source
Galbot (Chinese embodied AI startup) responds to comparisons of its Hong Kong retail stores to vending machines · embodied
Three Galbot Stores opened on September 1 (previously reported), selling bottled drinks, chips, and Hong Kong-themed souvenirs, quickly drawing comments that they're no different from vending machines. Founder and CTO Wang He responded that limited store space currently rules out coffee, ice cream, and hot food, and that the company will keep exploring product categories popular with Hong Kong shoppers. He said Galbot already operates over 200 stores on the Chinese mainland, with product mix adjusted to local demand at each location; in Hong Kong, the goal is to expand to dozens of stores, with site selection targeting high-foot-traffic areas like Disneyland and Central, and eventual expansion into pharmacies and tennis courts. On theft concerns, he said the robots record video throughout, so footage would serve as evidence if an incident occurred.Source: HK01 source
NOW Robotics wins KRW 2.4 billion automation equipment contract · industrial
The contract is with VPK, for equipment used on a hybrid vehicle battery pack production line at a US plant. The KRW 2.4 billion deal is modest in size, but points to the pattern of South Korean automation equipment makers expanding overseas alongside automakers' battery production capacity.Source: Asia Economy source
Industry Developments
Supreme People's Court issues AI-related judicial guidance: false advertising causing harm in autonomous driving brings civil liability · autonomy
At a September 7 press briefing, China's Supreme People's Court released its "Opinions on Lawfully Adjudicating AI-Related Disputes," comprising 5 sections and 24 provisions — the first judicial adjudication rules on AI issued by the country's highest court. Two vehicle-related provisions stand out. First, on liability allocation: the Opinions clarify liability rules for damage caused by autonomous vehicles and vehicles with driver-assistance features, noting that current cases concentrate on damage caused by assisted driving; where a vehicle defect combines with driver fault to cause the same harm, courts will support claimants who seek liability from both the driver and the manufacturer/seller. Second, on advertising: where vehicle manufacturers or sellers make false or misleading claims about automation level, intelligence, performance, or intended use, harming consumer interests, courts will support consumers' civil liability claims under the Civil Code and the Consumer Rights Protection Law. The Opinions also set out adjudication standards for AI face/voice swapping, AI "resurrection" of the deceased, algorithmic price discrimination, celebrity-impersonation livestream sales, doxxing, and open-source software liability. The focus on liability centers on assisted driving rather than L4 autonomy, consistent with the direction of the September 4 draft amendment to the Road Traffic Safety Law, which places liability for violations occurring while driving-assist features are active on automakers.Source: Sina News source
7 listed Chinese robotics companies' half-year reports: revenue doubling is common, profitability is the exception · humanoid
Revenue is up broadly across the board, but the picture turns once you reach net income. Among the 7 listed embodied AI companies tracked by The Paper, only the recently listed Unitree Robotics posted positive net profit attributable to shareholders: H1 revenue of RMB 1.152 billion, up 48.54% year-on-year, and net profit attributable to shareholders of RMB 274 million, versus a loss of RMB 32.02 million in the same period last year — though non-recurring-adjusted net profit was RMB 244 million, down 19.34% year-on-year, which the company attributed to larger increases in R&D and sales spending. UBTech (Chinese humanoid robotics company) posted revenue of RMB 1.269 billion, with a net loss attributable to shareholders of RMB 311 million, narrowing 24.7% year-on-year; H1 total humanoid robot sales reached 16,123 units, up 268.3% year-on-year, of which 921 units met the company's "full-size embodied AI humanoid robot" standard, contributing revenue of RMB 590 million — up from 6.1% of total revenue in the same period last year to 46.5%, surpassing the education business for the first time. Sales and shipment figures diverge sharply here: according to Smart Analytics Global, UBTech's H1 humanoid robot shipments were around 700 units. By the same source, Zhiyuan Robotics shipped around 8,400 units in H1, up 562% year-on-year, with global market share rising from 25% to 44% to lead the market; Unitree shipped around 5,900 units for a 31% share, ranking second. Unitree does not break out humanoid-specific revenue, disclosing only that as of July it had produced roughly 18,000 cumulative units across its various bipedal humanoid models. Luoshi Robotics (Chinese robotics arm maker) posted the fastest growth, with revenue of RMB 415 million, up 135.8% year-on-year, and adjusted net profit turning positive at RMB 18.4 million — though it's the core supplier of humanoid robot arms to Zhiyuan Robotics, so its growth is heavily concentrated in a single customer. Dobot (Chinese robotics company) posted revenue of RMB 316 million, up 106.6% year-on-year, but its loss actually widened to RMB 106 million, with R&D spending rising to 32.1% of revenue. Woan Robotics (Chinese robotics company) swung from profit to loss, with a net loss of RMB 36 million, R&D spending up 73.3% to RMB 102 million, and combined revenue of roughly RMB 68.4 million across three new embodied AI product lines. Estun Automation (Chinese industrial robotics company) narrowed its loss the most, with a net loss attributable to shareholders of RMB 56.9258 million, down 62.48% year-on-year, though its 15.32% gross margin was the lowest of the seven.Source: The Paper source
Two South Korean ministries jointly push physical AI for small factories, starting with AMR consulting for single-stage logistics · industrial
South Korea's Ministry of Science and ICT and Ministry of SMEs and Startups jointly visited KAIST's Physical AI verification lab on Sunday, convening technology suppliers and manufacturers with field needs, linking the science ministry's technology development and verification capabilities with the SME ministry's support policies. The two ministries are matching supply and demand through a Physical AI Alliance, with a three-phase roadmap: this year, consulting on introducing AMRs and AGVs for single-stage logistics; from 2027, support for connecting heterogeneous equipment across full logistics stages along with precision manufacturing data; and from 2028, expanding consulting and operating systems to integrated management of logistics and production. The technical foundation draws on physical AI full-stack technology developed through two verification projects — unmanned factory control and operations in Jeollabuk-do, and ultra-precision manufacturing intelligence in Gyeongsangnam-do. The SME ministry will handle environment-building and verification on the smart factory side, with successful practices to be folded into the integrated "ManufacturingAI 24" platform.Source: Digital Today source
Unitree claims first real-time world-model-driven fully autonomous humanoid combat · humanoid ⚠️ Manufacturer's own claim
Unitree Robotics announced on September 7 that UnifoLM-X2-1.0 achieves real-time world-model-driven fully autonomous humanoid robot combat, releasing footage alongside the announcement. The company describes this as breaking through bottlenecks in instantaneous planning, decision-making, and dynamic interactive execution for world-action large models. Unitree's humanoid robots have publicly demonstrated ring combat before — two units sparred on stage on the opening day of the China International Import Expo in November 2025 — but the company says the difference this time is that it no longer relies on preset routines or manual remote control.Source: IT Home source
Zhang Yiming personally leads ByteDance's real-time world model project · world-model ⚠️ Unverified/rumored
Bloomberg reported on September 7 that Zhang Yiming, who stepped back from day-to-day management years ago, is coordinating engineers across multiple business lines on a real-time world model built on top of the Seedance video model, with a release expected as early as next month. The report cites a target of roughly 50 milliseconds latency at 20 frames per second, rendered in the cloud and streamed to users, with a user's movements or voice able to alter the scene in real time. The target hardware is Pico, the VR headset maker ByteDance acquired for nearly $2.8 billion in 2021, with a parallel track aimed at TikTok's creator tools, keeping heavy compute on servers so creators don't need their own GPUs. The world model ranks first among ByteDance's four AI priorities for 2026, ahead of Seedance's video lead, coding tools, and Doubao's (ByteDance's AI chatbot) monetization. A data vendor told 36Kr that ByteDance's spending on world-model training data this year is three to four times that of its Chinese peers, in the eight-figure RMB range.Source: Startup Fortune, citing Bloomberg source
Zhixiang Future (Chinese AI startup) releases embodied world model, claims top spot on RoboColiseum's disturbance-adaptation leaderboard · world-model ⚠️ Manufacturer's own claim
HiDream-O1-Embodied made its debut on the embodied simulation benchmark platform RoboColiseum, with the company saying it ranked first on the Robustness leaderboard with an average score of 0.692. The platform sets 78 simulated evaluation tasks across four dimensions — instruction following, spatial understanding, disturbance adaptation, and general manipulation — with disturbance adaptation testing generalization by varying background, lighting, materials, robot starting states, camera position, and image quality. The data approach pairs "real motion-capture base + generative augmentation," using high-precision motion-capture data from Noitom (Chinese motion-capture company) as a foundation, then generating variant videos that preserve physical constraints while altering only background lighting and object shape. Zhixiang Future CTO Yao Ting called this a key milestone in moving from simulated to real-world environments. The company released its interactive world model, HiDream-O1-World, less than a month earlier.Source: Zhidongxi source
Zeno Robotics (Chinese robotics startup) releases 3B-parameter collaborative foundation model Zeno-1 · world-model ⚠️ Manufacturer's own claim
At 3B parameters, running local closed-loop visual-motor inference at 30 Hz, the company says it's the first physical intelligence foundation model designed specifically for decentralized multi-robot collaboration. In demonstrations, two robots hang up pants and vacuum-seal pillows using copies of the same policy model, with who grasps, who assists, and when to wait or take over all determined by each robot's own observations, with no central controller assigning roles or synchronizing clocks. Training proceeds in four stages, the key one being closed-loop partner interaction: two independently running robots act as each other's training partners, each robot's actions altering the shared physical environment and thereby influencing the other's next decision; since the partner is also learning and making mistakes during training, the model is exposed to a wide range of partner behavioral deviations. The technical report has been made public.Source: Sina Finance source
Oversized robot lithium batteries want to fly, rental operators say they have a workaround · adjacent
The Civil Aviation Administration of China's rules are clear: batteries rated at ≤100Wh can be carried on board without approval, 100Wh to 160Wh requires airline approval, and above 160Wh is banned both as carry-on and checked baggage. Unitree's Go1 standard model has a battery rated at 133.2Wh, and batteries for the rest of its quadruped and humanoid robot lineup all fail to meet requirements for personal air travel; Unitree customer service told National Business Daily that "it currently cannot fly," and shipping by air under a company's name requires special logistics certification. An industry source who asked to be identified by the pseudonym Xu Min said a fully packaged robot can only be shipped by land, and even shipping a spare battery alone gets refused by logistics companies even with a full set of dangerous-goods certification documents; one Hangzhou-based robotics startup shipped by sea for an overseas performance instead. Meanwhile, reporters posing as overseas customers contacted several robot rental operators, all of whom said they had "a way" and that "shipping the robot by air is no problem" — one Shanghai-based operator said it goes through the ATA Carnet channel, brings the battery along, and works with the airline to add weight allowance. An ATA Carnet is a customs document set up by the World Customs Organization for goods temporarily entering or leaving a country; it only resolves customs clearance and does not exempt shipments from air dangerous-goods regulations. A Hubei-based rental operator that handles Unitree robot business said it previously relied mainly on the ATA channel, but recently conditions have changed and it has temporarily stopped shipping robots together with their batteries by air.Source: National Business Daily source
Hardware & Supply Chain
· Dexterous hands: GGII data shows China's dexterous hand sales reached about 19,200 units in 2025, up 236.84% year-on-year, projected to hit 70,200 units by 2026 and potentially exceed 430,000 by 2030 — but the high-DOF five-fingered hands on display are not the same thing as what's actually selling well; shipments are still dominated by three-fingered grippers and low-DOF products. Fu Yuancheng of Xino Future (Chinese dexterous hand startup) says once degrees of freedom exceed 25, marginal benefit declines while marginal cost rises; around 20 active degrees of freedom already covers over 90% of human hand function, with motors accounting for about 20% of a dexterous hand's cost; a high-DOF dexterous hand retails from around RMB 100,000 to several hundred thousand, occasionally over a million. Ni Hualiang, CEO of Aoyi Technology (Chinese prosthetics/dexterous hand company), said the company has nearly 300 partners, of which 90% are still stuck at proof-of-concept, with 10% of customers contributing over 90% of revenue source
· The software ceiling for dexterous hands: Zhang Zhengtao, founder of Zhongke Huiling (Chinese robotics company), says no VLA model to date can drive a 20-to-30-DOF hand — "it can move, but lacks dexterity"; Fu Yuancheng added that current VLAs can output only around 10 motion-control vectors at a given timestamp, meaning they can only manage around 10 degrees of freedom of hardware at once — the industry has reached a point where software is the bottleneck constraining hardware source
· Rockchip chips: Microduck, a desktop duck robot from a French-American team, has sparked a wave of clones in China, driving shortages and price increases for Rockchip chips source
· Huaxing Yuanchuang (Chinese test equipment maker): humanoid robot controller module testing equipment has now shipped to overseas laboratories source
Top comments (0)