Today's Highlights
· Waymo filed a recall with NHTSA for nearly 4,000 robotaxis, yet not a single vehicle was taken off the road
· Lotte Mart's Busan center goes live, with 1,000 Ocado robots serving 4 million households
· Lens Technology (Chinese electronics manufacturer) teams up with CASIA Vision to launch a new company, unveiling three dexterous hands at once
· Kong Tao, former head of ByteDance's robotics team, now leads Xiaomi's robotics foundation model team
· ω-0 achieves 81.8% success across 11 real-world household tasks, nearly double the next-best baseline
· A Silicon Valley-adjacent county proposes a licensing scheme for humanoid robots, requiring emergency stop switches and continuous human supervision
· Moore Threads (Chinese GPU maker) posts H1 revenue of RMB 1.736 billion, up 147% YoY, and simultaneously announces plans for a Hong Kong listing
1. Research Progress
ω-0: a latent-space world-action model letting humanoid robots move and manipulate at the same time · locomotion
Household chores rarely offer the clean segmentation of "walk into position, then reach" — humanoid robots must shift their footing, adjust posture, maintain balance, and manipulate objects all at once. Rather than splitting the task into a navigation module plus an arm-control module, ω-0 predicts a whole-body action latent directly from language instructions, visual observations, and proprioceptive state, which is then handed to the controller for execution; the key design choice is to not generate future video, instead learning a compact future-observation embedding that supplies cues about task progress and scene evolution, with whole-body actions then generated via diffusion. Across 11 real-world household tasks, ω-0 achieves an 81.8% success rate, compared with 27.3% for π-0.5 and 22.7% for GR00T-N1.7; the strongest baseline, ψ-0, reaches only 44.5%. The team also released the ω-HOME dataset, over 40 hours of multi-view observations with whole-body action annotations.
Zhe Li et al. (MARS Lab NTU / Peking University / BAAI / HKUST Guangzhou) · arXiv 2608.06375 source
The "data pyramid" for embodied manipulation: a data-recipe survey co-authored by 11 institutions · benchmark
The field widely agrees that data is the bottleneck for embodied manipulation, but there has never been consensus on what mix of data is needed. This survey, jointly produced by 11 institutions, arranges real-robot data, UMI-style handheld capture, first-/third-person video, simulated data, and general vision-language data into a pyramid, with the central axis being the tradeoff between scalability and robot-alignment — the higher up the pyramid, the closer to real-robot data and the harder it is to scale. The survey further uses this data-recipe lens to review recent embodied foundation models, breaking down how each selects, aligns, and blends these sources during pretraining, and maps data composition onto five capability categories: perception, reasoning, planning, action generation, and world prediction.
arXiv 2607.24744 source · Analysis: blog.csdn.net source
2. Funding & Deals
Moore Threads (Chinese GPU maker) | H1 report + planned H-share listing | Revenue RMB 1.736 billion · hardware
Chinese GPU maker Moore Threads disclosed both its H1 report and Hong Kong listing plans on the same day, August 9: H1 revenue was RMB 1.736 billion, up 147.42% YoY, with net loss attributable to shareholders narrowing to RMB 11.5631 million (RMB 259 million less loss YoY, an 95.73% narrowing), and gross margin at 56.95%. The company attributed growth mainly to accelerated commercialization of its Kuae intelligent-computing clusters, naming embodied AI as one of the core scenarios for its compute infrastructure; its ultra-large-scale training cluster based on the MTT S5000 achieved 95% linear scaling, while on the edge side it launched the MTTAICUBE and E300 modules based on its "Changjiang" SoC. R&D spending was RMB 769 million, or 44.30% of revenue. Details of the Hong Kong share offering have not yet been finalized.Source: STCN (Chinese financial news outlet) source
Feiji Kesi | Series A1 | hundreds of millions of RMB · world-model
Existing investor TEDA Kechuang co-led the round with Hisound Capital and Fudan Kechuang, joined by new investors Yida, Youshi, and Yongyue, with existing investors Yunqi and Silicon Harbor increasing their stakes. Founded in 2024, the company builds physical-intelligence infrastructure — its self-developed differentiable physics engine Fysics, an embodied simulation training platform, and physical world models. It sells the "training environment and data" layer rather than robot hardware itself.Source: STCN Venture Connect source
Paxini Tech | Strategic round | RMB 1 billion | valuation over RMB 10 billion · hardware
Following disclosure of this round earlier this month, the investor list and cumulative funding total were filled in this week: the round was co-led by a consumer electronics and semiconductor giant, BOC International Investment, Kunpeng Fund, and Hexin Fangce, with participation from CIC Capital (Baifu Fund) and the Chengdu Jiaozi AI Fund, and existing investor Zhilai Capital increased its stake substantially. The company's cumulative funding has now reached RMB 3.5 billion; it focuses on the full chain from Hall-effect multi-dimensional tactile sensors to dexterous hands and complete humanoid robots, and has filed over 200 patents.Source: STCN Venture Connect source
Weekly primary-market roundup | RMB 8.459 billion disclosed | down 73.17% week-on-week · adjacent
From August 1–7, China saw 150 investment/financing deals (versus 147 the prior week), but disclosed funding fell from RMB 31.533 billion to RMB 8.459 billion — the deal count barely moved while the disclosed amount was more than halved twice over, indicating last week's high total was driven by a handful of large transactions. Advanced manufacturing led with RMB 4.164 billion, and embodied AI remained among the most sought-after sub-sectors; by funding stage, seed/angel rounds accounted for 33% of deals, while Series A led in total amount with RMB 2.506 billion.Source: STCN Venture Connect source
3. Commercialization & Deployment
Lotte Mart's Busan Zetta Smart Center goes live, with 1,000 Ocado robots taking over fresh-food fulfillment · industrial
This is South Korea's first customer fulfillment center to adopt the Ocado Smart Platform, and the first CFC built since Lotte Mart's partnership with Ocado (announced November 2022) with an investment of roughly KRW 200 billion: construction began in December 2023, was completed last August, and after integration testing was completed this July, the center began full-load operations. Across roughly 40,000 square meters, up to 1,000 picking robots ("BOT") travel across a grid at 4 meters/second, with storage stacked 8 to 21 levels high; robotic arms ("OGRP") pick from above the grid, adjusting grip strength via sensors to avoid crushing fresh produce. The chilled zone is kept at 5°C and the frozen zone at -25°C, with robots operating continuously. The operator says the entire facility is staffed by just over 300 people, with robots handling most processes aside from inbound stocking and final packing. Serving 4 million households across Busan and the Yeongnam region, daily delivery slots have increased from about 3 to as many as 13, with 2–3 hour delivery windows available, and an SKU ceiling of 35,000 — roughly 10,000 more than a typical store.Source: Chosunbiz, Asia Economy source
Ligong Industrial's full-size humanoid "Lidian" enters mass application in aerospace and auto-parts manufacturing · humanoid
This company, a nationally recognized "little giant" specialist enterprise with nearly 40 years of precision-manufacturing experience, has developed its own humanoid robot, "Lidian," with 28 degrees of freedom across its whole body and a peak joint torque of 320 N·m, and it has already achieved mass application at leading firms in aerospace and auto-parts manufacturing. Against this year's industry backdrop, where "over 70% of humanoids remain confined to showrooms and labs," a company with roots in parts manufacturing validating the robot on its own production lines first represents a different path from that of standalone humanoid startups.Source: Sohu.com source
Saite Intelligent's medical logistics robots cover over 95% of hospital material-transfer scenarios · industrial
Guangzhou-based Saite Intelligent's "Zhisaila" series now covers over 95% of hospital material-transfer scenarios, having cumulatively served over 1 million patients; its "Zhisaijie" autonomous sweeping vehicles have been deployed at Guangzhou Baiyun railway station, the Baiyun International Convention Center, and multiple stadiums, capable of centimeter-precision edge cleaning. The company holds over 700 proprietary IP rights, and R&D outcomes from the provincial-level medical service robot engineering center it established have generated over RMB 300 million in commercialized value. In-hospital medical logistics is one of the few service-robot scenarios that has already scaled, where success hinges on regulatory compliance and dispatch reliability rather than robot form factor.Source: Sohu.com source
Robot dogs start taking over security guard shifts, at $80,000–$130,000 less per shift annually than human labor · adjacent
According to Business Insider, AI camera-equipped quadruped robots are being used to patrol data centers, farmland, and stadiums, covering a 24/7 post at $80,000 to $130,000 less per year than hiring a human. Asylon, a robot-dog company, said its CEO reports about 50 robots currently deployed across 25 clients — still a tiny fraction relative to the 1.3 million security guards employed nationwide in the US. Researchers at UC Berkeley's Labor Center note that the more realistic near-term effect isn't replacement, but pushing human workers toward the harder, higher-risk portion of the job; the occupation's 2024 median annual salary was $38,390.Source: Futurism source
4. Industry Developments
Waymo files recall for nearly 4,000 robotaxis, yet not a single vehicle stops operating · autonomy
Waymo has filed a recall with NHTSA covering nearly 4,000 robotaxis, over a defect in which its driving software poorly recognized highway construction zones and ramp closures, repeatedly driving into areas that should have been closed off in San Francisco and Phoenix. Waymo's own phrasing was that it "identified room for improvement in highway construction-zone performance" — in substance, the vehicles didn't know that road conditions had changed since the last mapping update. Two signals stand out here. First, self-driving has spent a decade proving it can handle pedestrians, cyclists, and unprotected left turns, yet traffic cones plus a single flagger turn out to be harder — construction zones are scenarios that deliberately break the prior mapping assumptions the system relies on, and standardized, digitized real-time lane-closure feedback from construction crews is still unevenly adopted across the US. Second, the recall framework dating back to the 1966 Motor Vehicle Safety Act presumes an "owner": a letter is mailed, the owner visits a dealer, a part is replaced. Waymo owns and operates its own fleet, so notifying owners means notifying itself, and the fix is simply pushing a software update — meaning not a single vehicle was taken offline, and no rider even noticed; one outlet even had to issue a correction after initially reporting that the vehicles had been taken out of service. Regulatory vocabulary is being stretched to cover a fleet that can be patched overnight. Waymo has not yet opened highway driving in Texas, so riders there were not exposed to this defect.Source: The Auto Wire source
Kong Tao, former head of ByteDance robotics, now leads Xiaomi's robotics foundation model team · world-model
Multiple independent sources confirm that Kong Tao, former head of ByteDance's robotics team, now leads Xiaomi's robotics foundation model team, having joined in summer 2025 along with a number of former ByteDance colleagues. Xiaomi's robotics division now has about 200 people, working across hardware, motor control ("cerebellum"), cognition ("brain"), and even robot operating systems, with different teams sometimes running parallel research on the same direction (such as world models); the foundation model team has its own separate office space. Kong Tao received his PhD from Tsinghua University's computer science department in 2019 and was ByteDance's first employee focused specifically on robotics research; his personal homepage states that as of June 2025 his robotics-related papers had been cited over 15,000 times, and he left ByteDance in June 2024. Xiaomi released Xiaomi-Robotics-0 in February and Xiaomi-Robotics-1 in July this year, with the former sharing the "brain + cerebellum" paradigm used in GR-3, which Kong Tao worked on during his time at ByteDance. Talent flow hasn't been one-directional either: a number of people from Xiaomi's CyberOne and CyberDog 2 teams later joined ByteDance Seed, filling exactly the hardware and motion-control gaps there.Source: 21st Century Business Herald source
Lens Technology (Chinese electronics manufacturer) teams up with CASIA Vision to found Zhongke Huisi, launching three dexterous hands · hardware
Zhongke Huisi Embodied Intelligence (Hunan) was jointly founded in July 2026 by Beijing CASIA Vision, Lens Technology, and Hunan Huaxia Investment Group, and made its debut in Changsha on August 7 with the launch of three dexterous hands — L1, D1, and M1: the L1 takes a lightweight, reduced-DOF approach with 6 active degrees of freedom at a price in the thousands-of-RMB range, positioned as an upgrade replacement for existing grippers, and has already been used to have the CASBOT BAND robot band play guitar, bass, and electronic keyboard; the D1 is a high-DOF general-purpose five-finger hand dedicated to real-robot data collection; the M1 uses a modular "standardized body + functional fingertips" architecture aimed at industrial scenarios. The company also pre-announced a tendon-driven, bio-inspired F-series. Based in Hunan's Xiangjiang New Area, the company says it will build the country's largest dexterous-hand skill training ground and data-collection center. Its strategy is a dual-flywheel model of "multiple robot platforms × data collection" — first lowering the barrier to actually putting these end-effector "hands" to use through a series of hardware platforms, then feeding the resulting data back to train skill models.Source: Sohu.com, STCN source
Silicon Valley-adjacent county proposes licensing scheme for humanoid robots · humanoid
San Mateo County Supervisor Ray Mueller will submit a proposal to the county Board of Supervisors next week requiring a licensing process for autonomous/semi-autonomous humanoid robots deployed in commercial settings such as restaurants and retail. Cited reasons include the need for specialized firefighting equipment against battery thermal-runaway fires, unpredictable robot behavior potentially obstructing evacuations, and pressure on the tax base and social services from job losses in the service sector. Applicants would need to demonstrate that their robots comply with ADA and health standards, are equipped with an emergency stop switch, have no outstanding recalls or history of catastrophic failure, and are continuously supervised on-site by trained personnel during operation; licensing fees would go toward firefighting equipment procurement. Mueller expects state or federal legislation to eventually take over the matter, but argues the county should act first. This is among the first local initiatives to treat humanoid robots as a distinct regulatory subject, following a period when regulatory attention was concentrated mainly on self-driving cars.Source: Palo Alto Daily Post source
BYD's first bipedal humanoid "Xiaodi" debuts in Zhengzhou this month, entering through dealerships rather than factories · humanoid
BYD confirmed it will showcase its first bipedal humanoid robot, "Xiaodi," this month at the Di Space brand experience center in Zhengzhou, where it will interact with customers and introduce vehicle models. BYD Executive Vice President and Americas CEO Stella Li said the goal is "two or three units per dealership," and believes robot sales assistants could become commercially viable within one to two years. According to media reports, the robot stands 1.61 meters tall, weighs 58.5 kg, and has 31 degrees of freedom (7 per arm, 6 per leg), with 1mm hand positioning precision and a 1kg payload — specs better suited to a controlled showroom environment than industrial work.⚠️ Manufacturer's own claimsCompared with Hyundai's approach of betting on mass-producing Atlas at a US factory by 2028, BYD is prioritizing retail service before scaling toward industrial deployment — read by some as reflecting that Chinese factory-floor scenarios are already claimed by players like Unitree, UBTECH, and AgiBot.Source: The Korea Herald source
Turnover in AgiBot's partner roster ahead of IPO; questions remain over Luo Jianlan's status · embodied
Following earlier reports this week that AgiBot's (Chinese humanoid robot startup) official partner roster grew to 9 people, with Chief Scientist Luo Jianlan absent from the list, more of the timeline has since been filled in: on July 24, the same day AgiBot confirmed it had initiated its Hong Kong listing process, Shanghai's Institute for Advanced Study in AI released a video stating that Luo Jianlan had joined the institute as a full-time mentor; references to his AgiBot role on both his personal homepage and X bio have since been removed. AgiBot's response was that he has "not yet left; this is an internal role adjustment," without disclosing a new title.⚠️ Single-party accountThe three newly added partners — Xiong Yan, Zhu Jie, and Zhang Xiuzheng — come from Huawei's Central Research Institute, Tencent Public Affairs, and Huawei Cloud's China region respectively, all with industry rather than research backgrounds; together with founder Deng Taihua, nearly half of the 9 partners now have Huawei backgrounds. Luo Jianlan, a student of Sergey Levine, led the open-sourcing during his tenure of AgiBot World, a dataset of over 1 million real-robot manipulation trajectories, along with the τ0-WM and τ0-VLA models, and at June's Beijing Academy of Artificial Intelligence conference bluntly stated that "chasing leaderboards is meaningless."Source: Caijing Tianxia WEEKLY source
MirroS proposes Physical RSI: moving recursive self-improvement from code sandboxes into the physical world · world-model
Recursive self-improvement (RSI) in the digital world — where an AI edits its own code, runs experiments, and retains changes that work, in a closed loop — carries an implicit assumption: that the environment is known and fixed. MirroS's proposed Physical RSI argues that in the physical world, this must instead be a dual-axis coupling, with the Actor and World Model iterating in tandem: when something unexpected happens, the system first checks whether its world model can reproduce the causal chain of the failure (an "Environment Gap"); only once it can does the system move on to searching for a new strategy (a "Skill Gap"). A related claim is that world models should be built from language abstractions plus executable code rather than end-to-end latent representations, on the grounds that when a black-box model fails, its errors can't be localized, making "improvement" meaningless.⚠️ Single-party accountThis is currently a framework and research roadmap, with no public experimental results yet.Source: Sohu.com source
Guangzhou Baiyun industrial robotics cluster named a provincial specialty industrial cluster · industrial
The cluster currently comprises 75 companies, spanning robot bodies, core components, system integration, and AI large-model and application services; 2025 output value exceeded RMB 24.7 billion, with an average annual growth rate of 18.85% over the past three years, and exports of leading products rose from RMB 3.789 billion in 2024 to RMB 6.34 billion. The cluster includes 1 nationally recognized single-item manufacturing champion and 11 nationally recognized "little giant" specialist enterprises. In 2025, the cluster's R&D spending was RMB 1.238 billion, an R&D intensity of 5.01%, with local supply-chain coverage rising to 73%.Source: Sohu.com source
Hardware & Supply Chain
· Dexterous hand shipments: According to GGII data, China's dexterous hand sales in 2025 were approximately 19,200 units, up 236.84% YoY, projected to reach 70,200 units in 2026, up 265.63%; CASIA Vision's Zhang Zhengtao estimates the industry is growing at roughly 3.7x annually based on this trajectory source
· Dexterous hand cost and lifespan: Zhang Zhengtao says market prices for high-DOF tactile dexterous hands range from the low hundreds of thousands of RMB to RMB 800,000–1,000,000+, and that "Tesla's Optimus hand still costs $6,000," but under continuous industrial use, actual lifespan is only a few weeks to two or three months source
· "Hands without brains" and the data gap: He also notes that no VLA model to date can genuinely drive a 20–30 DOF hand to do real work; open-source data for complete robot systems already includes millions of trajectories, while open-source data for fine dexterous-hand manipulation "amounts to only tens of thousands — several orders of magnitude behind" source
· Foundation Robotics' tendon-driven hand: Motors are positioned in the forearm and pull tendons to open/close each joint, keeping the fingers slim and fast-moving; finger position is inferred from motor rotation angle and cable routing, with magnetic encoders at each joint serving as an angular backup — the hand has already demonstrated catching a thrown baseball in mid-air, though the throw's trajectory was pre-planned and there is no tactile feedback yet; it will be used in the company's factory-oriented Phantom humanoid source
Top comments (0)