Today's Highlights
· Ant Group's Robbyant open-sources LingBot-VLA 2.0, using 60,000 hours of data to let a single brain drive 20 different robot embodiments
· Momenta lists on the Hong Kong Stock Exchange as the "first Physical AI stock," raising about $750 million, shares up slightly on debut
· World-model papers surge on the same day: Alibaba's RynnWorld-4D HF↑70, though the industry is debating whether "world model" has become a catch-all buzzword
· LimX Dynamics completes a Pre-IPO round of nearly $200 million, signaling clear IPO intent for its humanoid robots
· Waymo extends robotaxi service to 4 new cities including San Diego, expanding its driverless footprint
I. Research Papers
RynnWorld-4D: A 4D Embodied World Model for Robotic Manipulation · world-model
Today's highest-trending world-model paper on Hugging Face (HF↑70). It argues that RGB alone is insufficient for manipulation, and that RGB, depth, and optical flow must be jointly modeled (RGB-DF) to capture the scene's 4D dynamics under interaction — letting a robot not just "recognize what a scene looks like" but also predict how its 3D structure evolves as actions unfold. From an Alibaba team.
Haoyu Zhao et al. (Alibaba) · arXiv 2607.06559 source
RynnWorld-Teleop: "Digital Teleoperation" via World Models for Data Generation · world-model
HF↑65. Robot data collection has long been bottlenecked by physical teleoperation — every demonstration ties an operator's time to specific hardware. This work proposes replacing real robots with a generative world model, decoupling data collection from physical constraints and offering a cheaper path to generating data for scalable robot learning. From the same Alibaba team as RynnWorld-4D.
Haoyu Zhao et al. (Alibaba) · arXiv 2607.06558 source
AlayaWorld: Long-Horizon, Playable Video World Generation · world-model
HF↑64. Aimed at games and interactive scenarios, this work replaces the traditionally labor-heavy world-building pipeline with a video world model, emphasizing "long-horizon + playable" generation — meaning the generated world can be continuously interacted with by a player rather than drifting apart or collapsing after a few seconds.
AlayaWorld Team, Kaipeng Zhang et al. · arXiv 2607.06291 source
LingBot-VLA 2.0: One VLA Brain for 20 Robot Embodiments · vla
An open-sourced general-purpose VLA foundation model (weights and code included) from Ant Group's Robbyant, dominating attention across both industry and community today. It rebuilds the data pipeline with 60,000 hours of data (50,000 hours of real-robot trajectories spanning 20 embodiments, plus 10,000 hours of human first-person video), supports the full body — head, waist, mobile base, and dexterous hands — via a 55-dimensional unified action space, and closes the "predicting the future" temporal gap with a token-level loss-free MoE action expert plus dual-query distillation, directly targeting the industry's chronic problem of "dazzling in the lab, falling apart on real hardware." HF↑10.
Wei Wu et al. (Ant Group Robbyant) · arXiv 2607.06403 source · Analysis: The Way of Embodiment source (WeChat, CN)
World Models' Long-Horizon Failures Are Really "Kinematics Without Dynamics" · world-model
A diagnostic paper offering a more specific explanation than "error accumulation" for why world models' imagined futures grow increasingly implausible the longer they run: models tend to imagine the future based on kinematics (how objects should move) rather than dynamics (how forces evolve), causing them to gradually drift from physical reality over long-horizon rollouts. A sobering mechanistic breakdown amid the current overheated world-model hype.
Finn Rasmus Schäfer et al. · arXiv 2607.05966 source
FORGE: Letting Robots "Grab Whatever Tool Is at Hand," Like Humans Do · manipulation
Humans can casually grab a book, a rock, or a shoe to use as a hammer, but robots trained on specific tools fail when given a novel object. FORGE formalizes this "functional generalization" by using keypoint trajectory reasoning to let a policy focus on "how a tool is used" rather than "what a tool looks like," enabling the same function to transfer to unseen objects.
Chuhao Zhou et al. · arXiv 2607.05780 source
Putting the Arms on the Shins: A New Quadruped Dual-Arm Loco-Manipulation Design · locomotion
Most quadruped manipulation designs force a tradeoff between "manipulation capability" and "stance stability." This work integrates dual arms onto the shins of a quadruped robot, enabling two-handed tasks without needing to rear up or lean back — offering a new embodiment design that balances payload and stability for quadruped mobile manipulation.
Yan Pan et al. · arXiv 2607.06186 source
Other papers today: MoWorld, MoxinCore's Flash world model, 50FPS with roughly 70% lower cost (arXiv 2607.06216 source); A Definition and Roadmap for World Models, clarifying definitions and a roadmap for the overheated "world model" term (arXiv 2607.06401 source); Learning 4D Geometric Priors for Inference-Efficient World Action Models (arXiv 2607.05468 source); Training-Free Acceleration for VLA, training-free inference speedup for VLA (arXiv 2607.06370 source); Optimal Transport Q-Learning, accelerating flow-policy inference and guidance (arXiv 2607.06262 source); LAMP, latent action priors driving real-robot dexterous-hand learning (arXiv 2607.06323 source); ThorArena, a humanoid physical-interaction benchmark (arXiv 2607.06052 source); Lift3D-VLA, adding 3D geometric and dynamics reasoning to VLA (arXiv 2607.06564 source); Why does Deep Learning Improve Visual SLAM? (Scaramuzza's group, arXiv 2607.06023 source).
Open Source · Tools · Benchmarks
· RoboDojo (the embodied "college entrance exam"): A new benchmark pits 30 mainstream robot policies against each other, covering 42 simulated and 18 real-robot tasks; the results are sobering — humans score 100, the strongest model scores only 12.8, laying bare the current true state of embodied capability. source (WeChat, CN)
· TAP (open-sourced by Fudan University & Shanghai Innovation Institute): Lets robots first learn "how to move," then "how to follow instructions," decoupling motor skill from language alignment — training motor skills first actually raises success rates. source (WeChat, CN)
· Palette NEAT (open-sourced by SiMa.ai): An open-source, edge-focused tool that accelerates deployment of Physical AI models. source
· ForesightSafety-VLA: The first unified safety-diagnostic benchmark for VLA, defining 13 categories of safety risk and revealing the widespread problem of current VLAs "appearing successful while actually being dangerous." source (WeChat, CN)
II. Funding & Deals
LimX Dynamics | Pre-IPO round | ~$200 million · humanoid
A Shenzhen-based full-size general-purpose humanoid robot company, following its $200 million Series B in February this year (with participation from JD.com, SAIC Motor, NIO, and others), has now raised nearly $200 million in a Pre-IPO round, with clear intent to pursue a listing. The company got its start in legged locomotion control, with a product line spanning full-size humanoids and its multi-form robot TRON.Source: guandian.com source
SIASUN Duoco Robotics (part of the Chinese Academy of Sciences' SIASUN group) | Series A+ | Hundreds of millions of yuan · industrial
Jointly led by Hongtai Fund and CDH Baifu, with follow-on participation from Chengtong Fund, Volkswagen JuDing, and others. The company is an industrial embodied-AI enterprise under the SIASUN group of the Chinese Academy of Sciences, with a product lineup covering full humanoid units, force-controlled robotic arms, and core joint modules — its force-controlled robotic arms have already shipped in volume in 2026. Unlike pure-algorithm teams, its moat lies in over a decade of industrial accumulation.Source: Tech Capital Circle source (WeChat, CN)
Wujie Power | Angel round | Over $200 million · embodied
Jointly invested by a JD.com-affiliated fund, C Capital, and Hony Capital, with a separate Pre-A round of nearly $200 million nearing completion — bringing cumulative fundraising in H1 2026 to several hundred million dollars. The company focuses on general-purpose embodied manipulation intelligence and a "general-purpose brain," founded by Zhang Yufeng, co-founder of Megvii (Chinese computer-vision company).Source: Stark on the Future source (WeChat, CN)
Queue | New round | $12.6 million · adjacent ⚠️ Single-source account
The lead investor was drawn to the critical automation infrastructure the company has built for pharmacies. The company describes itself as building "the world's first fully automated robotic pharmacy," using robotics and automation to ease labor shortages in the pharmacy industry and improve prescription-filling and pickup efficiency.Source: Economic Digest source (WeChat, CN)
Dogtooth (UK) | New round | £14 million · embodied
A UK agricultural harvesting robotics company, with the funding aimed at helping growers cope with labor shortages. Automated picking of soft fruit (such as strawberries) is a recognized hard problem in agricultural automation, requiring both ripeness recognition and compliant/flexible grasping to be solved simultaneously.Source: Business Matters source
III. Commercialization & Deployment
Waymo Robotaxis Launch Simultaneously in 4 New Cities · autonomy
Waymo is rolling out driverless rides simultaneously in San Diego, Las Vegas, Tampa, and Denver, starting in a "riders limited to Waymo employees only" mode for now, with the company saying public access will follow soon; its new vehicle model, the Hyundai Ioniq 5, is currently being tested with a safety driver on board. Following Tesla's Robotaxi launch in Miami (previously reported), the leading players' driverless footprints continue to expand.Source: The Verge source
CATL-Backed Autonomous Mining Truck Company Easun Zhijia Lists on Hong Kong Stock Exchange · autonomy
Easun Zhijia (Eacon), an L4 autonomous mining-truck company backed by CATL, listed on the Hong Kong Stock Exchange today, opening up about 9%, rising as much as 14% intraday, with a market cap of about RMB 12 billion (about HK$13.5 billion). Headquartered in Shanghang, Fujian, the company specializes in driverless mining operations and already runs mining trucks at scale in real mine sites; however, its prospectus shows cumulative losses of about RMB 1.2 billion over three years, meaning the path from scaled operations to profitability still has some way to go.Source: 36Kr source
AgiBot (Chinese embodied-AI startup) Demonstration Project Launches in Shanghai, Deploying with Aptiv Factory · embodied
Embodied-AI company AgiBot announced a partnership with global auto-parts giant Aptiv, planning to deploy 100 robots at Aptiv's factory in 2026 — billed as "the country's first hundred-unit-scale industrial embodied-AI cluster" — with plans to scale to a thousand units by 2027. ⚠️ Planned/demonstration-stage figures; the target unit counts still need to be verified through actual unit-by-unit delivery.Source: Shanghai Economy and Information Technology Commission source (WeChat, CN)
Midea KUKA Signs with Winner Medical, Production-Line Automation Orders Exceed RMB 300 Million · industrial
Midea KUKA signed agreements with Winner Medical and Winner Recycling, with a deal scope exceeding RMB 300 million, providing automation and smart-manufacturing upgrade solutions for their medical-consumables and recycled-materials production lines. Large-scale production-line automation orders driven by manufacturing leaders like this remain the most reliable current source of cash flow for industrial robotics.Source: guandian.com source
Amazon Unveils Next-Generation Warehouse Robots, Adds $12 Billion European Expansion · industrial
Amazon unveiled its next-generation AI warehouse robots and announced a $12 billion European expansion plan. The new robots feature enhanced autonomous package recognition and hand-off capability, pointing toward "handoff-free" package flow within warehouses. As one of the world's largest buyers of warehouse automation, its model iterations often set the pace for industry-wide deployment.Source: AOL source
New Hope Group Begins Raising Pigs with Robots · industrial
Amid a sharp drop in pork prices and squeezed farming margins, New Hope Group (Chinese agribusiness conglomerate) is replacing manual labor with automation to cut costs, pushing embodied/automation technology deeper into large-scale agriculture.Source: Sina Finance source
IV. Industry Developments
Momenta Lists on Hong Kong Stock Exchange, Dubbed the Market's "First Physical AI Stock" · autonomy
Autonomous-driving company Momenta (stock code 6880.HK) rang the bell on the Hong Kong Stock Exchange today. The global offering raised about $750 million, priced at the top of the indicative range, after public subscription had at one point been oversubscribed by about 414 times (previously reported); shares rose modestly by about 3% on their trading debut. Momenta started out in autonomous driving and in recent years has packaged its R7 world model together with real-world driving data into a "data flywheel" narrative to raise capital, with founder Cao Xudong comparing mass-produced autonomous driving to "building a rocket capable of mass production before going to the moon." As "world model" hype heats up across the industry, this IPO is being read as a market test of whether the "world model + mass-production gateway" story can actually deliver.Source: 36Kr source (WeChat, CN)
"World Model" Is Being Criticized as AI's Latest Catch-All Buzzword · world-model
With Alibaba's RynnWorld-4D, MoxinCore's MoWorld, GigaAI's GigaWorld, and other world-model releases all landing on the same day, "world model" is increasingly being questioned as a term at risk of becoming AI's newest "catch-all buzzword." Wang Zhongyuan, president of the Beijing Academy of Artificial Intelligence (BAAI), divides the space into four categories — language, pixel, 3D, and visual representation — and stresses that "video generation ≠ world model": generating realistic pixels has never been the same as reproducing the laws of physics. Fei-Fei Li, Yann LeCun, and Jensen Huang have recently disagreed publicly over the definition and direction of world models, and several papers published today (including "A Definition and Roadmap for World Models") are likewise attempting to pin down the boundaries of a term that capital is chasing hard.Source: 36Kr source
Huawei "Genius Youth" Alumni Cluster Into Embodied-AI Startups; Zhao Lichen Targets an "Embodied Agentic OS" · embodied
According to 36Kr, nine companies founded by former Huawei "Genius Youth" program alumni have together raised over RMB 7.7 billion. Among them, Zhao Lichen, a former class-of-19 "Genius Youth" member, left to start a company targeting an "embodied Agentic OS"; the company has raised over RMB 1 billion just six months after founding, with the team transplanting the data-flywheel and mass-production quality-control systems proven in autonomous driving wholesale into embodied-AI R&D, aiming to shorten the cycle from technical validation to mass production.Source: Sina source
Mistral AI Releases a Robotics Model, Doubling Down on Physical AI · world-model
French frontier AI company Mistral AI has released a robotics model, formally entering Physical AI. According to Bloomberg, the company recently acquired Austrian industrial physics-simulation firm Emmi AI, integrating physics simulation, engineering knowledge, and robotics capability into its enterprise platform for manufacturing, aerospace, energy, semiconductor, and other engineering use cases. This marks another European entrant into the embodied/Physical AI race, following Nvidia and several Chinese foundation-model makers.Source: Bloomberg source
BYD Officially Enters Humanoid Robotics, First Use Case Targets "Smart Retail Assistance" · humanoid
BYD has officially announced its entry into humanoid robotics, disclosing that its first deployment scenario will target "smart in-store shopping assistance." As one of the world's largest EV makers by shipment volume, BYD holds advantages in motor, motor-control, and vertically integrated supply-chain capabilities; it currently remains at the scenario-planning and initial-launch stage.⚠️ Planned-stage figuresSource: Autohome source
Reconova (Chinese computer-vision company) Lists on Hong Kong Stock Exchange, "First Visual Embodied-Intelligence Stock" Falls Below Offer Price on Debut · autonomy
Reconova, listed on the Hong Kong Stock Exchange and dubbed the "first visual embodied-intelligence stock," fell below its offer price on debut, closing down about 3.05%, with a market cap of about HK$5.5 billion. Amid soaring embodied-AI valuations in the primary market, secondary-market investors are pricing the "embodied AI" label considerably more cautiously.Source: The Paper source
Hardware · Supply Chain
· Dexterous hands: Q1 funding in this segment reached nearly RMB 5 billion, with full-year shipment volume projected to exceed 70,000 units; Chinese dexterous hands have already achieved 0.1mm-level precision, with single-unit lifespans surpassing 1 million cycles. source (WeChat, CN)source (WeChat, CN)
· OFILM: Its machine-vision business now operates independently and has entered the supply chains of robotics companies including UBTech and DEEP Robotics. source
· New MIT chip: Builds real-time 3D maps for robots using only the power draw of a single LED, pointing toward lower-power perception for vehicles and robots. source
· BIOSTAR EdgeComp MS-NAT4000: A compact edge-AI computing system built on the NVIDIA Jetson Thor T4000, aimed at onboard compute for robot embodiments. source
Top comments (0)