Today's Highlights
· A modified Unitree G1 robot remotely completed the world's first live laparoscopic cholecystectomy (gallbladder removal), published in Nature
· XPeng Robotaxi begins employee testing, with He Xiaopeng himself taking the first ride — just 8 months from announcement to a vision-only L4 pickup
· Ant Group's LingBot open-sources LingBot-World 2.0, described as the world's first hour-scale real-time interactive world model
· Mistral releases its first robotics model, Robostral Navigate, which navigates from a single camera following spoken English instructions
· Wujie Dongli's (Chinese industrial humanoid startup) K15 wins the world's first industrial-grade full-scope CE certification, launching near-$100M in order deliveries
1. Research Papers
LaMem-VLA: Compressing History into "Latent Memory" for Direct Use in VLA Reasoning · vla
Mainstream VLA (vision-language-action) models operate under a Markov assumption, seeing only the current observation — a real handicap on long-horizon tasks that require remembering earlier context. Existing memory-augmented approaches either bloat the observation window or bolt on a retrieval store, both of which slow reaction time. LaMem-VLA reconstructs history into a set of "latent memory" tokens that interleave directly with the reasoning chain, letting memory guide action generation within a bounded context. With HF↑41, it's the most upvoted robotics paper in this batch.
Hongyu Qu et al. · arXiv 2607.07608 source
LingBot-Video: First Open-Source MoE Video Foundation Model Built for Embodiment, Not Aesthetics · world-model
General-purpose video generation models prioritize visual fidelity and creativity, which is a poor fit for physical realism and compute efficiency — using them directly for robot control tends to break down. Ant Group's LingBot adapted the DiT + MoE video pretraining paradigm into an embodiment-focused version, releasing it as the first large-scale open-source MoE video foundation model, aiming to give "robot brains" a video engine capable of reasoning about physical consequences. HF↑37.
Shuailei Ma et al. (Ant Group LingBot) · arXiv 2607.07675 source · Coverage: Infoq.cn source (WeChat, CN)
LingBot-World 2.0: An Unbounded-Duration, Self-Evolving Interactive World Model · world-model
Previous interactive world models were limited by generation duration and consistency, drifting and breaking down after tens of seconds. This paper (corresponding to the open-sourced LingBot-World 2.0) achieves unbounded interaction duration without quality collapse via causal pretraining, and is the first to bring "agentic" orchestration into world modeling — a pilot agent plans character behavior while a director agent synthesizes new scene elements on the fly, letting the virtual world "grow" autonomously. Now open-sourced, HF↑19.
Zelin Gao et al. (Ant Group LingBot) · arXiv 2607.07534 source · Coverage: Techgogogo source (WeChat, CN)
TouchWorld: A Foundation Model for Manipulation Where Touch Both Predicts and Reacts · perception
Dexterous manipulation requires both anticipating how contact will evolve and correcting slippage, misalignment, or unstable grips within milliseconds — vision alone can't reach this level. TouchWorld treats touch simultaneously as a "predictive contact reference" and a "fast feedback signal," significantly improving local contact adaptation while preserving the semantic generalization of VLA models — a rare attempt to elevate touch to the foundation-model level.
Jianyi Zhou et al. · arXiv 2607.07287 source
NativeMEM: Native Long-Horizon Memory, Matching Prior Work with a Fifth of the Training Data · vla
Long-horizon manipulation needs frequently updated visual history, but external memory management typically forces a tradeoff between "memory length" and "policy responsiveness." NativeMEM builds long-term, real-time-updating memory directly into the policy itself, consistently outperforming prior methods while using only 20% of the training data — data efficiency is its biggest selling point.
Ziye Wang et al. · arXiv 2607.06678 source
WAM-TTT: Correcting World-Action Models at Test Time by "Watching Humans" · world-model
Steering robot foundation models toward new task variants or user preferences usually requires additional robot demonstrations or dedicated fine-tuning. WAM-TTT instead uses test-time training: raw human videos are absorbed via self-supervised video prediction into a lightweight adaptive memory inside a frozen world-action model, then a meta-training stage aligns human demonstrations with robot behavior — enabling adaptation "by watching humans" without new real-robot data.
Yusen Feng et al. · arXiv 2607.06988 source
Pelican-VLA 0.5: Without Any Labels, the Action Pathway Learns to Attend to the "Right" Objects on Its Own · vla
Unifies vision-language understanding, future-frame generation, and action prediction in a single architecture. The standout result is its generalized "attention-level" behavior: with no object labels, segmentation masks, attention supervision, or task-specific fine-tuning, its action pathway spontaneously focuses on instruction-relevant objects and contact regions — hinting at a more label-efficient path to generalization.
Zeyuan Ding et al. · arXiv 2607.06655 source
Ace!: A Robotic Arm Serves Table Tennis at Professional Level, Spin Up to 550 rad/s · manipulation
Table tennis has long been a classic benchmark for dynamic robot control, but most research has focused on returning shots rather than serving. This paper uses motion primitives + model predictive control + Bayesian optimization to generate legal serves, achieving controllable spin up to 550 rad/s and ball speeds up to 6.7 m/s — matching or exceeding human players in spin and speed, a nice result for agile manipulation.
Guillem Torrente et al. · arXiv 2607.06989 source
Other papers today: SPEAR, a photorealistic embodied simulation library that can programmatically control arbitrary Unreal Engine applications (4 citations, arXiv 2607.06701 source); EmbodiedGen V2, an agentic simulation-ready 3D world engine (arXiv 2607.07459 source); GeoProp, a plug-in adapter for geometric alignment between proprioception and vision (arXiv 2607.07101 source); RoboSnap, generating a simulatable scene from a single RGB image (arXiv 2607.06699 source); a survey of VLA models for drones and dual-arm manipulation (arXiv 2607.06706 source); WildCity, a city-scale real-world rendering and spatial intelligence testbed (HF↑4, arXiv 2607.06838 source); spatial-relation grounding and "instruction leakage" fixes for compact world models (arXiv 2607.06925 source); Flow-ERD, traffic simulation balancing realism and diversity, currently topping the WOSAC leaderboard (arXiv 2607.06957 source); CILC, cryptographically secure loop closure detection for multi-robot collaborative SLAM (arXiv 2607.06700 source); GemNav, discrete-token vision navigation using a frozen multimodal LLM plus LoRA (arXiv 2607.06882 source).
Open Source · Tools · Benchmarks
· LA4VLA: Open-sourced by Shanghai Jiao Tong University and Alibaba, arguing for learning language-action relationships separately — temporarily removing visual input during training for language-action pretraining, to prevent VLA models from taking the shortcut of "acting from images" without learning stable language-action constraints source (WeChat, CN)
· Robo-ValueRL: Open-sourced by Beijing Humanoid Robot Innovation Center (Chinese humanoid research consortium), adding a "data quality discrimination" capability to embodied models so robots no longer indiscriminately absorb all training data — targeting the data-quality bottleneck for scaling source (WeChat, CN)
2. Funding & Deals
AI²Robotics (Zhipingfang) | Series Funding | Nearly RMB 5 billion (~$700M) | Valuation over RMB 20 billion · embodied
Dubbed the "most Tesla-like" general-purpose embodied AI company, this round made it the first embodied AI unicorn in the Guangdong-Hong Kong-Macao Greater Bay Area valued over RMB 20 billion. Investors unusually span state-backed funds, government ministries, Bay Area platforms, insurers, brokerages, and multiple companies in Tesla's supply chain ecosystem. Technically the company pursues full in-house development along a single end-to-end VLA path, and in June released its open-source brain-inspired model NeuroVLA (a cortex-cerebellum-spinal-cord hierarchy that cut robotic arm motion jerk by over 75% and reflex response time to under 50ms). Sources: PEdaily source (WeChat, CN) · Huxiu source
Qiuzhi Technology (Wuxi) | Angel Round | Over $100M · embodied
Incubated by Tsinghua's Institute for AI Industry Research (AIR), backed by two major investors including Lenovo, setting a record for the largest angel round in China's consumer-grade embodied AI sector. Funds will go toward core technology R&D, prototype-to-mass-production, and home service scenarios; the company's proprietary DISCOVERSE simulation platform forms the base of an algorithm-model-hardware-training full stack. A nine-figure angel round signals capital moving early on the home embodied AI sector.Sources: Xinhua Daily Finance source (WeChat, CN) · Chuangke Gongshe source (WeChat, CN)
General Intuition | New Round | $320M | $2.3B valuation · world-model ⚠️ Single-source claim
Positioning itself as the "foundation layer of physical AI for robotics," the company is betting on a counterintuitive thesis: using video game data instead of real-robot telemetry to train embodied AI's "GPT moment." It claims its model can play games continuously for hours and, after fine-tuning on just 8 minutes of real-robot data, drive a quadruped robot — self-reported experimental claims awaiting third-party replication. The company doesn't build robots itself; it sells a general-purpose behavior model.Sources: RoboEra source (WeChat, CN) · Silicon Canals source
3Srobotics (Sheng Shi Wei Sheng) | Series B | Hundreds of millions of RMB · industrial
An embodied AI company focused on industrial manufacturing, this round led by Shanghai Semiconductor Industry Investment Fund and Jinqiao Fund, with 1Vision Capital and Xinding Capital participating. The company's flagship product is a proprietary "welding embodied brain" that puts an intelligent welder on the production line — another example of Shanghai state capital piling into industrial embodied AI.Sources: Robot Qianzhan source (WeChat, CN) · 36Kr source
DeepCtrls | Series B | Hundreds of millions of RMB · embodied
A company positioned around "physical AI" completed a Series B round of hundreds of millions of RMB, led by a state-backed fund with Sequoia participating. State capital paired with a top-tier USD fund is a typical configuration in today's physical AI primary market: state backing plus market-driven upside.Sources: Zhidongxi source · Pandaily source
Stord | Series F | $250M | $3B valuation · adjacent
Builds a "physical intelligence layer" for e-commerce fulfillment, integrating warehousing, transportation, and order flow into a single hardware-software network. Its Series F of $250M at a $3B valuation shows the "physical intelligence" narrative spilling over from robot hardware into supply-chain software.Source: aol.com source
3. Commercialization & Deployment
Wujie Dongli's K15 Wins World's First Industrial-Grade Full-Scope CE Certification, Launches Near-$100M in Order Deliveries · industrial ⚠️ Order figures self-reported
Wujie Dongli's (Chinese industrial humanoid startup) K15 became the world's first embodied AI robot to obtain industrial-grade full-scope CE certification — CE is a third-party compliance gate for EU market access, and clearing it means legal entry into European industrial settings, a genuine milestone. The company simultaneously announced it is beginning global mass delivery of "near-$100M in orders" — order size and delivery progress are company-disclosed and not yet independently verified. Following an over-$200M angel round the previous day, the company has quickly pivoted focus to overseas compliance and mass-production delivery.Sources: PEdaily source · Dahe Caili Fang source
Yinshi Robotics' Changshu Smart Manufacturing Base Opens, Dexterous Hand Deliveries Topped 10,000 Units in 2025 · embodied ⚠️ Market share figures self-reported
Yinshi Robotics (Chinese dexterous-hand maker), which focuses on humanoid five-finger dexterous hands, formally opened its Changshu smart manufacturing base, with its bionic five-finger hand now at its seventh generation. The company says dexterous hand shipments exceeded 10,000 units in 2025, with global market share over 60%, serving both Chinese and international leading humanoid and EV manufacturing clients — the shipment and market-share figures are company-reported. Moving from prototypes to deliveries "in the tens of thousands" is a rare scaling signal in the dexterous-hand segment.Sources: Yinshi Robotics source (WeChat, CN) · Changshu National High-tech Zone source (WeChat, CN)
LinkerBot's Changping Dexterous Hand Smart Production Line Launches, Planned Annual Capacity of 60,000-100,000 Units · embodied ⚠️ Capacity is a planning target
LinkerBot (Chinese dexterous-hand maker) established a wholly-owned subsidiary, "Lingdong Zhixin," in Changping, Beijing, building a smart production line with planned annual capacity of 60,000-100,000 dexterous hands, witnessed during an on-site inspection by Beijing municipal leadership. The capacity figure is a planning target, not yet validated through production ramp-up. The company is also reportedly filing for a Hong Kong Stock Exchange IPO, with market estimates putting its valuation above RMB 100 billion. The move toward a production line for "robots building robots" marks a key step in shifting dexterous hands from manual assembly to mass production.Sources: Industry Toutiao source (WeChat, CN) · Gasgoo Community source
Hanjin Launches South Korea's First Autonomous Delivery Truck Service · autonomy
South Korean logistics giant Hanjin launched South Korea's first autonomous delivery truck service, extending driverless technology from robotaxi passenger service into trunk/last-mile freight. An East Asian logistics firm entering L4 freight through a genuine commercial service rather than road testing offers a new benchmark for deployment pacing.Source: 매일경제 source
Midea's KUKA Signs Another Order Over RMB 300 Million in Foshan · industrial
Midea-owned KUKA landed another industrial robot order over RMB 300 million in Shunde, Foshan; in the same period Midea's Welling Robotics component manufacturing base signed a new agreement, planned to begin production in 2027. A home appliance giant is combining its own manufacturing use cases with component production capacity to advance industrial automation on two fronts.Sources: OFweek source · Gasgoo source
4. Industry Developments
XPeng Robotaxi Begins Employee Testing, He Xiaopeng Becomes "Passenger No. 1" · autonomy ⚠️ Employee-testing stage
XPeng held its first all-hands meeting for the Robotaxi business and launched employee testing, with founder He Xiaopeng personally completing the first ride, running the full chain from online booking to automated pickup to drop-off. The vehicle is based on the flagship GX model, described as China's first vision-only, fully in-house-developed, factory-installed mass-production Robotaxi, equipped with 4 proprietary Turing AI chips (3,000 TOPS of local compute) and the Turing VLA 2.0 model, with L4 capability; the "world's highest" compute claim is company-sourced. The business is positioned as a "product and technology provider," partnering with platforms like Amap for operations, with operating licenses and accident liability resting with the platform. Just 8 months from last November's announcement to employee testing outpaces industry expectations, though employee testing is not commercial operation, and the regulatory and liability questions around vision-only L4 remain unresolved.Sources: Shenzhen Commercial Daily source (WeChat, CN) · Cyzone source (WeChat, CN)
Mistral Releases Its First Robotics Model, Robostral Navigate, Navigating from a Single Camera on Spoken English Commands · world-model ⚠️ Company-sourced claims
Leading European AI lab Mistral makes its first move into robotics with Robostral Navigate, an 8B-parameter model that navigates offices, homes, and outdoor environments using natural-language instructions and just one ordinary RGB camera, with no depth sensor. The model was trained entirely in simulation (roughly 400,000 paths across 6,000 virtual spaces), representing actions as "pointing" to target pixel coordinates plus local displacement. The company self-reports a 76.6% success rate on the unseen R2R-CE test set, 4.5 points above the best multi-sensor approach — the benchmark figures are company-disclosed. Following collaborations with BMW and Airbus, Mistral is entering industrial automation with a lower-sensor-cost approach.Sources: Reuters source · Bloomberg
Yuanli Lingji Releases DM0.5 General-Purpose Embodied Model: 4B Parameters, 150,000 Hours of Data · embodied ⚠️ Company-sourced claims
Yuanli Lingji (Chinese embodied AI startup) unveiled five new products at once in a "three-stage rocket" launch, centered on DM0.5, a 4B-parameter general-purpose embodied model trained on 150,000 hours of data, emphasizing zero-shot capability, long memory, and robustness for real-world deployment. The capability claims and "emergent generalization" language come from the launch event and have not been independently benchmarked. Amid the day's broader narrative of "data scaling," it shifts the competitive axis from parameter count to accumulated real-world interaction hours.Sources: Jiqizhixin source (WeChat, CN) · Ifeng source
Rokae Lists on Hong Kong Stock Exchange, Raises Roughly HK$810M Net in Global Offering · industrial
Rokae (Chinese industrial robotics maker), founded 11 years ago after 10 funding rounds, listed on the Hong Kong Stock Exchange, billing itself as the "first full-line smart robotics stock," with a global offering raising roughly HK$810 million net. Its products serve industrial scenarios including 3C electronics and auto parts. Alongside same-day listings from RVBUST Technology and Momenta, the Hong Kong listing window for robotics and autonomous driving companies remains hot.Sources: Beijing News source · EBRun source
AgiBot Signs Strategic Partnership with ASD (Chinese cookware maker), Focused on Live On-Site Training · industrial
AgiBot (Chinese humanoid robotics startup) signed a comprehensive strategic partnership with cookware manufacturer ASD, and held a "deployment-mode" industry sharing event with an on-site live training exhibition zone, pointing toward embodied robots moving from prototypes into routine commercial use. According to Omdia, AgiBot held roughly 39% global market share in humanoid robot units in 2025. The partnership uses a real manufacturing production line as its deployment entry point.Sources: Gaogong Robotics source (WeChat, CN) · Jiqi Fengchao source (WeChat, CN)
Mitsubishi Motors Teams with a Startup to Enter Humanoid Robotics, Plans Manufacturing and Deployment by 2027 · humanoid ⚠️ Planning stage
Mitsubishi Motors announced a partnership with a startup to develop humanoid robots, planning to manufacture and deploy "humanoid workers" in its own facilities by 2027. This is a planning-stage announcement with no confirmed mass-production timeline yet delivered. Following Toyota and Honda, another Japanese automaker enters the field, with humanoid manufacturing becoming a second front for automakers.Source: Automotive News source
Hardware & Supply Chain
· RoboSense: Shipped 719,200 LiDAR units in H1 2026, with its robotics segment up 510.4% year-over-year — beyond automotive, robotics is becoming a new growth driver for LiDAR source
· LG Electronics' "Axium" humanoid joint modules: Began early mass production of "Axium," a joint module for humanoid robots, as the Korean appliance and components giant stakes out a position in humanoid actuator supply chains source
5. This Week's Watch
NHTSA Upgrades from "Guidance" to "Order": Robotaxis Must Not Obstruct Emergency Response · autonomy
U.S. NHTSA Administrator Jonathan Morrison sent a letter to the industry explicitly stating that emergency scenarios "are not rare edge cases" for driverless vehicles, and that they must not interfere with emergency vehicles or first-responder operations — a shift in regulatory tone from industry guidance to an enforcement-level directive, seen as tightening the "right of way" for robotaxi scale-up. At the same time, U.S. regulators are loosening rules elsewhere: according to Bloomberg, the U.S. is exploring allowing driverless vehicles without a steering wheel on public roads. One tightening, one loosening — the L4 regulatory framework is still taking shape.Sources: The Drive source · Luobo Taikeqi source (WeChat, CN)
China's Ministry of Industry and Information Technology and SASAC Launch Special Initiative on Live On-Site Training for Humanoid Robots and Embodied AI · embodied
China's Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission (SASAC) jointly launched a 2026 special initiative in June, targeting over 100 high-value application scenarios and enabling deployment "at the scale of tens of thousands of units" by year-end, focused on on-site training spaces and innovation application consortiums. This is a government action rather than a company claim — even with forward-looking targets, it constitutes real policy steering for China's embodied AI mass-production timeline. Local governments are moving in step: Shenzhen released its "Embodied Intelligence Robotics Technology Innovation and Industrial Development Action Plan (2025-2027 Revision)," and China's Ministry of Human Resources and Social Security is planning to establish a new job classification, "Embodied Intelligence Robot Application Technician."Sources: Industry Big Data Products source (WeChat, CN) · Embodied Intelligence Robots & AI source (WeChat, CN)
Citi: Physical AI Has Reached Commercialization, But Scaling Is a "Decade-Long Marathon" · adjacent
Citi's AI summit poured some cold water on the current enthusiasm: physical AI has crossed the commercialization threshold, but constrained by data scarcity and high costs, scaling robotics will be a "decade-long marathon" rather than a one-or-two-year breakout. This single-analyst view stands in contrast to the industry narrative of a "year of mass production," a reminder to separate company "tens-of-thousands-of-units" targets from real unit economics.Sources: Proactive source · Futu Niuniu source
This Week's Supply Chain Roundup: Reducer Localization Leads the Ramp-Up, Ball Screws Remain the Deepest Bottleneck · hardware ⚠️ Figures include company/analyst claims
This week's in-depth hardware coverage centered on a hot-and-cold divergence between two components. Harmonic reducers are now leading the front of "Chinese localization + production ramp-up": Green Harmonic Drive (Chinese reducer maker) raised capacity from 50,000 units/month at the start of the year to 70,000 units/month, planning 100,000-120,000 units/month by year-end, with order visibility extending to 2027; the company's market share among China-made harmonic-reducer-equipped humanoid robots is reported to have risen from about 70% to 80-90% (company-disclosed). SKF taking a stake in a joint venture with Green Harmonic Drive further embeds foreign precision-bearing expertise into China-made joints. The global robotic harmonic reducer market is estimated at roughly $8.7 billion in 2025, projected to exceed $14.8 billion by 2030.
By contrast, planetary roller screws were repeatedly flagged this week as the segment with the "deepest chokepoint": a single humanoid robot requires roughly 14 units, with per-unit value of about RMB 9,000 (nearly 19% of hardware cost, the highest of any component), while Chinese localization rates are widely cited at under 30%, with more aggressive estimates as low as 8-20%; core grinding machines remain dominated by Japanese and German suppliers, with delivery lead times of 1-2 years. Wuzhou Xinchun (a supplier to Optimus Gen3), Top Group (which secured an exclusive Tesla contract), and Hengli Hydraulics are among those racing for the "zero-to-one" opening in Chinese localization. In short: the next major battleground for cutting whole-robot costs is shifting from rotary joint components to linear transmission components.Sources: Wenya Liaoba source (WeChat, CN) · Ligen Consulting source (WeChat, CN) · Niuniu Haoniuniu source (WeChat, CN)
Top comments (0)