Today's Highlights
· Rokae Robotics listed on the Hong Kong Stock Exchange, closing up 15.2% on its first day, becoming the "first full-range intelligent robotics stock"
· China's mandatory national standard for autonomous driving enters public comment; implementation in 2027 will require L3 vehicles to have heterogeneous redundant sensing
· Unitree Robotics sets IPO price at RMB 46.80, aiming to raise RMB 4.5 billion, with subscription opening July 14
· Inspire Robots (Chinese dexterous-hand maker) ramps annual dexterous-hand capacity to 50,000+ units, holding over 60% global market share, with mass production at its Changshu base
· Tsinghua's Harness VLA wraps frozen policies without fine-tuning, boosting cross-distribution manipulation by 38.6 percentage points
1. Research Progress
Harness VLA: Wrapping frozen VLAs into "retryable contact primitives," boosting cross-distribution manipulation by 38.6 points · vla
End-to-end VLAs tend to collapse on long-horizon trajectories the moment a task needs redirection, target re-binding, or recovery from a single failed contact — and re-fine-tuning is expensive. A joint Tsinghua-led team took a different approach: without touching a single parameter of the VLA, they wrap the whole model into VLA_ACT, a "retryable" contact primitive, paired with a small fixed set of move/rotate/gripper parsing primitives, leaving "when and how to invoke" to an upper-level agent equipped with task memory and global rules. On the perturbation benchmark LIBERO-Pro, this pushed the previous best method from 43.8% to 82.4% (+38.6pp), while standard LIBERO performance held nearly unchanged at 96.0%, and long-horizon kitchen tasks on RoboCasa365 gained another 25.4pp. The core contribution is a ready-made, training-free path for extending frozen policies.
Tsinghua University · Purdue · Institute of Automation, CAS · HKUST et al. · Analysis: 具身智能漫谈 source (WeChat, CN)
TouchWorld: Building touch into a robot foundation model so "it looked right but didn't actually press" stops happening · manipulation
A camera can't tell the difference between "finger resting on a button" and "button actually pressed" — touch can. TouchWorld, a tactile foundation model from Yang Shuo's team at Harbin Institute of Technology, Shenzhen (Phanes AI, Chinese startup), makes touch do two jobs at once: before an action, it predicts a future tactile map of "what the contact should feel like"; after contact, it rapidly corrects course using real-time pressure/joint feedback. The key design choice is splitting the system across three time scales — 1Hz planning / 10Hz execution / 30Hz tactile correction — since tactile signals are too sparse to be fed directly into a unified VLA without being drowned out by vision. Across six real-robot tasks including watering plants and plugging in outlets, average success rate reached 65.0% (57.2% with added human-induced perturbations), beating the strongest baselines by 15.7 and 16.0 points respectively; with memory augmentation, the 4B planning model reached 91% accuracy, surpassing a 32B zero-shot model.
Yang Shuo et al. (Harbin Institute of Technology, Shenzhen / Phanes AI) · Analysis: 机器之心 source (WeChat, CN)
How should "world models" even be evaluated? A Nanjing University team publishes a position paper to set a ruler for an overused concept · world-model
The term "world model" today can refer to dynamics modeling, future video generation, interactive environments, and other quite different technical objects, leaving evaluation standards muddled. A Nanjing University team has published a position paper on evaluation, systematically mapping the concept's capability boundaries and evaluation dimensions in an attempt to untangle this catch-all term — a rare return to methodological groundwork amid a boom in world-model funding and demos.
Nanjing University · Analysis: 机器之心 source (WeChat, CN)
RoboBench: A cognitive checkup for the embodied "brain" · benchmark
A robot completing a task doesn't mean it truly "understands" the task. RoboBench is designed to evaluate the cognitive capabilities of embodied brains, testing a model's understanding of a task rather than just whether the outcome was achieved — offering a quantifiable diagnostic for the gap between "can execute" and "can reason."
Analysis: 具身智能之心 source (WeChat, CN)
VLA-Corrector: A lightweight 40M-parameter module patches VLA's "open-loop blind spot," learning to detect and correct its own errors · vla
Most mainstream VLAs execute open-loop and struggle to notice their own mistakes. A team from Zhejiang University and Alibaba DAMO Academy built a correction module with just 40M parameters that lets a VLA detect deviations during execution and correct them on the fly, adding closed-loop error correction to end-to-end policies at minimal cost.
Zhejiang University · Alibaba DAMO Academy · Analysis: 具身智能之心 TechDaily source (WeChat, CN)
ANY3D-VLA: Unified training across three types of point clouds eases 3D data scarcity and depth-scale bias · vla
A team from the University of Hong Kong and Peking University unified simulator-generated point clouds, real sensor point clouds, and depth-estimation-model point clouds into a single training pipeline, with 2D/3D feature alignment, to ease three long-standing problems: scarcity of embodied 3D data, cross-environment variation, and depth-scale bias.
HKU × Peking University · ICML 2026 · Analysis: Ru-hulu 具身智能分享会 source (WeChat, CN)
Other papers today: CoRE-VLA (conditional routing of experts, keeps running even with missing sensors, analysis source (WeChat, CN)); LaMem-VLA (dual-scale latent memory cures long-horizon tasks' "goldfish memory," Nanjing Univ. of Science and Technology / Zhejiang Univ. / NUS, source (WeChat, CN)); DexVerse (dexterous manipulation benchmark, highlighting that "can grasp" is still far from "can use," source (WeChat, CN)).
Open Source · Tools · Benchmarks
· LeRobot new release: Hugging Face's open-source framework update adds support for models and hardware including Fast-WAM and LingBot-VA source (WeChat, CN)
· NVIDIA Isaac GR00T: released a full end-to-end humanoid policy development pipeline — Isaac Lab-Arena simulation, Teleop demonstration collection, GR00T 1.7 post-training, and Jetson Thor deployment, along with a companion RoboLab policy evaluation suite source (WeChat, CN)
· Rice University × NASA: jointly open-sourced a remote space-robotics simulator for on-orbit manipulation policy training and evaluation source
2. Funding & Deals
Rokae Robotics | Listed on the HKEX Main Board | Net proceeds of approximately HK$810 million | Market cap approximately HK$11.68 billion · embodied
Listed on the Hong Kong Stock Exchange on July 9 at an offer price of HK$38, closing at HK$43.78 on its first day for a 15.2% gain, becoming the Hong Kong market's "first full-range intelligent robotics stock." Subscriptions were red-hot, with the Hong Kong public offering 156.58x oversubscribed; if the over-allotment option is fully exercised, total proceeds could reach approximately HK$1.006 billion. Founded in 2014 starting from controllers, the company is China's only manufacturer to mass-produce both industrial robots and collaborative robots, and by 2025 sales volume it was China's largest supplier of flexible collaborative robots, with a 47.0% market share. Financially it is still narrowing losses: 2025 revenue was RMB 522 million, with adjusted net loss narrowing to RMB 41.67 million, and Q1 shipments exceeded 5,100 units (including over 2,000 embodied robots).Source: Sina Finance / 第一机器人 source
Agility Robotics | Planned SPAC listing | $2.5 billion valuation | $620 million raised · humanoid ⚠️ Not yet listed
U.S. bipedal humanoid leader Agility plans to go public on Nasdaq via a merger with Churchill Capital XI (CCXI), at a valuation of $2.5 billion — roughly 34x trailing revenue (approximately $37 million) — backed by Nvidia, Amazon, SoftBank Vision Fund 2, and Foxconn. Its Digit robot has already logged over 65,000 hours of commercial operation across 9 enterprise customers, with over $300 million in outstanding RaaS (robot-as-a-service) orders on the books. However, while the valuation is set, the stock is not yet tradable: the S-4 registration statement has not yet been filed, shareholder votes cannot proceed until it takes effect, and the bulk of the order backlog rests on a single 1,000-unit, three-year contract contingent on the Digit v5 passing safety certification.Source: Tech Times source
Estun Automation (002747) | Plans all-cash acquisition of 100% of Cloud Ranger Intelligent · industrial
Estun, which just took the top spot in China's industrial robot sales volume with a 10.6% market share, plans an all-cash acquisition of Nanjing Estun Cloud Ranger Intelligent (酷卓), a unit it has incubated for years, to fill its gap in humanoid robotics. Cloud Ranger builds collaborative and humanoid robots; its second-generation humanoid Codroid 02 has already launched, with 2025 shipments up 270% year-over-year to a cumulative total of over 3,200 units, and its humanoid has landed its first batch of dozens of orders and completed testing at JD.com. Cloud Ranger's 2024 pre-A round (RMB 130 million) implied a post-money valuation of approximately RMB 560 million, which could serve as a pricing anchor. The controversy is that both sides are still burning cash: Estun's 2025 revenue hit a record RMB 4.888 billion, yet net profit attributable to shareholders was under RMB 45 million (non-recurring profit just RMB 7.28 million), while Cloud Ranger has racked up roughly RMB 100 million in cumulative losses over three years and remains unprofitable — consolidation risks further dragging down the income statement.Source: NetEase / 深水财经社 source
Unitree Robotics | STAR Market IPO priced at RMB 46.80/share | Aiming to raise approximately RMB 4.5 billion · humanoid
Following its IPO registration taking effect on July 11, Unitree has set its offer price at RMB 46.80/share, implying an issuance market cap of approximately RMB 42 billion, with total planned proceeds of approximately RMB 4.5 billion; subscription opens on July 14. Separately, the company was recently selected by the China Securities Regulatory Commission for an on-site inspection (a routine regulatory procedure).Source: 开元机器人简讯 source (WeChat, CN)
Lumisition Robotics | Angel round | Nearly RMB 50 million · hardware
Led by Zhengxuan Capital, with participation from Guangdian, Lihe, Songhe, and Hengxin Future. Founded just in February 2026, this Shenzhen-based company was started by a CUHK PhD holder and former DJI engineer, targeting consumer-grade quadruped robots and all-terrain mobility chassis, with funds earmarked for product development and mass production. The recent cluster of quadruped funding rounds reflects that chassis/mobility platforms are being treated as a low-barrier entry point into consumer-grade embodied robotics.Source: 机器人开放社区 source (WeChat, CN)
Weitai Robotics | Pre-A round | Backed in succession by Xiaomi and Will Semiconductor · hardware
Founded just two years ago and calling itself "the Nvidia of touch," Weitai Robotics has closed a Pre-A round with two major industry investors, Xiaomi and Will Semiconductor, joining in, with the next round already nearly closed. The team has an MIT lab background and focuses on robotic tactile sensing — touch is becoming the next perception frontier that industrial capital is racing to stake out, following vision.Source: 商业切片 source (WeChat, CN)
3. Commercialization & Deployment
Inspire Robots' Changshu base goes live, ramping dexterous-hand annual capacity from 10,000 to 50,000+, holding over 60% global market share · hardware
On July 9, Inspire Robots' smart manufacturing base in the Changshu High-Tech Zone went into operation, with its seventh-generation five-finger humanoid dexterous hand and micro servo cylinder production lines running at full speed — dexterous-hand annual capacity jumped from the 10,000-unit tier to over 50,000 units, and micro servo cylinder capacity exceeded 500,000 units. Founded in 2016 as China's first company to achieve mass commercial production of five-finger dexterous hands, the firm develops its core components in-house end-to-end, including micro motors, precision reducers, and planetary roller screws. It delivered nearly 10,000 units in 2025 (5x the prior year), with customers including Unitree, AgiBot (Zhiyuan Robotics), CATL, and BYD, and is targeting shipments of 30,000–50,000 units in 2026. The dexterous hand is the bottleneck component for the "last centimeter" of humanoid manipulation, and the leader pushing capacity into the hundred-thousand-unit range is a key step in pulling forward both whole-machine costs and delivery timelines.Source: 皖苏汇 source (WeChat, CN)
RoboSense's robotics business revenue surges 510% in H1 · perception
Lidar maker RoboSense disclosed that its robotics business revenue surged 510% year-over-year in the first half, becoming its fastest-growing segment — a sign that upstream perception-component makers are genuinely capturing the benefit of embodied robotics scaling up, rather than merely fielding order intentions.Source: 第一电动网 source
KEENON's service robot shipments remain #1 globally, while showcasing its humanoid lineup · industrial
KEENON Robotics says its global service robot shipment volume continues to lead the market, while also showcasing its humanoid product line. Delivery and cleaning service robots remain among the few categories currently deployed at real scale, and sustained shipment leadership speaks more to a viable commercial loop than humanoid demos do.Source: TipRanks source
Tesla's Texas factory to launch Cybercab employee ride service · autonomy ⚠️ Plan-stage claim
Tesla announced it will launch a Cybercab employee ride service at its Texas factory, but operational details, scale, and safety data remain entirely undisclosed. The employee pilot is a capability demonstration rather than a scaled ridership metric, and its real significance won't be clear until it opens to the public and regulatory safety data is disclosed.Source: Sina Finance source
Unitree G1 live-surgery paper now officially published (previously reported) · embodied
On July 10, Nature officially published the paper on Unitree's G1 performing a live-pig gallbladder removal; several outlets covered it again this week with no new technical or deployment developments to report.Source: 土豆先生备忘录 source (WeChat, CN)
4. Industry Developments
China's mandatory national standard for autonomous driving enters public comment, effective 2027, pure-vision approaches run into a compliance wall · autonomy
China's mandatory national standard for autonomous driving has entered its public comment stage and is slated to take effect in 2027, explicitly requiring L3-level vehicles to have heterogeneous redundant sensing capability — a hard bar that pure-vision approaches will naturally struggle to clear, since their sensors share a common failure mode. Vehicle models relying primarily on pure vision may be unable to legally open L3 highway functions, blocking their commercialization path. The standard also introduces Safety Case documentation and a DSSAD "black box," requiring full-scenario raw data disclosure — a review process that multi-sensor solutions, with their centimeter-level point clouds, will find easier to pass. Combined with New Jersey's proposed legislation restricting pure-vision autonomous vehicles, global regulation is shifting from technology exploration toward safety-baseline constraints, with compliance capability becoming a third axis of competition after compute and algorithms. Manufacturers with multi-sensor production experience (Arcfox, Changan Deepal, and others have already received L3 approval) gain an edge, while Tesla's pure-vision Cybercab route faces uncertainty.Source: 网通社 source
XPeng releases second-generation VLA autonomous driving model · autonomy
XPeng has released its second-generation VLA (Vision-Language-Action) autonomous driving model, which industry watchers are calling an attempt to seize a "DeepSeek moment" for autonomous driving, strengthening scene understanding and decision coherence under an end-to-end architecture. As pure-vision and multi-sensor camps diverge over the new national standard, the pace of in-vehicle VLA iteration is becoming the core bet for the pure-vision camp continuing to double down on algorithms.Source: 汽车之家 source
Symbotic acquires ARMS, moving warehouse operations from automation to AI coordination · industrial
Warehouse automation firm Symbotic has acquired ARMS, expanding its capability from single-point automation equipment to warehouse-level AI orchestration — a move to shift from "single-machine automation" toward "whole-facility coordination."Source: NetEase source
Tesla removes Model S/X production lines in 46 days to make room for Optimus, while Robotaxi faces new regulatory inquiry · humanoid ⚠️ Single-source claim
According to reports, Tesla removed its Model S/X production line at its Fremont factory in just 46 days to free up capacity for Optimus, with a supply-chain target of reaching 1,000 units per week by September — currently still a capacity-planning statement, with no mass-production output data yet available. Separately, U.S. federal regulators (NHTSA) have raised new inquiries regarding Robotaxi safety.Source: driveteslacanada / Yahoo Finance source
Official data: robotics exports near RMB 20 billion in first 5 months, over 1.15 million robotics-related enterprises now registered in China · industrial
According to official figures, robotics exports reached nearly RMB 20 billion in the January–May period, while the number of registered robotics-related enterprises in China now exceeds 1.15 million, with both the industry's scale and its export volume climbing in tandem.Source: 机器人行业统计 source
Barclays: humanoid robot market could reach $200 billion in 10 years; Insta360 officially enters robotics · humanoid ⚠️ Research-note / statement claim
A Barclays research note claims "Automation 3.0" has begun, with the humanoid robot market potentially reaching $200 billion in ten years (a long-range estimate); separately, Insta360 has officially announced its entry into robotics, another example of a consumer imaging company crossing over into embodied AI.Source: Cailian Press / 第一电动网 source
Hardware · Supply Chain
· 1X NEO dexterous hand: a 25-degree-of-freedom tendon-driven hand with fingertip tactile sensing that measures both normal and shear force to detect object slippage; the company says it plans to produce 10,000 units within the year (⚠️ manufacturer's own capacity claim) source (WeChat, CN)
· UTStarcom Vision Technology (天准科技): its machine-vision inspection equipment won a National Science and Technology Progress Award (Second Class), another authoritative endorsement for the localization of upstream vision inspection in China source
Top comments (0)