Today's Highlights
· Nvidia launches Halos for Robotics, billed as the industry's first full-stack safety system for physical AI, extending automotive-grade safety architecture to robotics; Agility Robotics is the first to integrate.
· Bear Robotics acquires UK humanoid startup Kinisi Robotics, folding in the KR1 robot and the Bristol engineering team, expanding from "mobility + delivery" to an end-to-end physical AI platform that includes dexterous manipulation.
· Pony.ai (Chinese autonomous driving company) fully opens its Singapore Robotaxi service to the public, marking another overseas commercial milestone for China's autonomous driving sector.
· General Motors deploys roughly 50 robots at its Detroit Factory Zero plant after previously cutting more than 1,000 jobs, drawing fierce UAW protests as the automation-vs.-labor debate intensifies.
· Haiqing Zhiyuan (Chinese physical-AI sensor company) lists on the Hong Kong Stock Exchange, dubbed the "first physical AI IPO" by the market; shares surged more than 300% at the open on debut day.
I. Research Papers
Ψ₀: Open-source whole-body loco-manipulation VLA backbone for humanoids — new skills from just 80 real-robot demonstrations · vla
Extends the paradigm of "large-scale human-video pretraining + minimal real-robot fine-tuning" to full-body loco-manipulation on humanoids, and provides a complete sim-to-real deployment pipeline on the Unitree G1 — reproducible and immediately usable. The most deployment-ready paper today.
Commentary: Embodied AI Open-Source Repository source (WeChat, CN)
Ψ₀ first learns general manipulation priors from large-scale first- and third-person human video, then fine-tunes on only roughly 80 real-robot demonstrations per new skill to achieve transfer — covering coordinated mobile-base and upper-limb manipulation. The authors also release an end-to-end pipeline from simulation training to real-robot execution on the Unitree G1 (Chinese robotics company), lowering the barrier to reproducing humanoid loco-manipulation results.
DeMaVLA: First generalizable VLA for deformable objects — one model handles multi-category garment folding · manipulation
Deformable objects such as garments have long been among the hardest manipulation challenges — infinite shape variation makes category-specific policies difficult to generalize. This work attempts to complete multi-category folding with a single VLA backbone, tackling one of manipulation's hardest open problems.
Midea AIR-C team · Commentary: Embodied AI Observatory source (WeChat, CN)
Existing VLAs largely rely on category-specific policies and struggle to generalize across diverse deformable objects and scenes. DeMaVLA takes multi-category garment folding as its core challenge, training a single model to complete long-horizon "grasp → unfold → align → fold" sequences across garment types, with the goal of moving beyond the one-policy-per-category engineering paradigm.
GenHOI: Driving humanoid-object interaction from generated video all the way to real-robot execution · manipulation
Bridging "video generation" to "real-robot execution" is a critical step toward deploying world models. GenHOI aims to turn generated human-object interaction footage into executable robot actions — not just visually compelling video.
Commentary: Embodied AI Research Lab source (WeChat, CN)
The method first generates video of a humanoid interacting with objects, then parses and retargets the interaction trajectories into executable robot actions, completing the full pipeline from generated footage to real-robot manipulation — exploring a "video as data / video as planning" approach. Specific success rates and generalization scope should be verified against the original paper.
UDHM / UniDexTok: A shared language for dexterous hands — from five fingers to six to twenty-four · manipulation
Dexterous hands vary widely in morphology and their training data is not interoperable, which is a root cause of difficulty in sharing and reusing manipulation policies. This work maps human hands and multiple robot hands into a unified joint-representation space, providing a foundation for cross-morphology transfer.
Commentary: Embodied Habitat source (WeChat, CN)
The Unified Dexterous Hand Model (UDHM) maps the joint poses of human hands and multiple robot hands into a shared 22-dimensional active-joint coordinate space, with dimensions defined as semantic joints based on human hand anatomy. This allows hands with different degrees of freedom to be expressed, transferred, and aligned under the same representation, facilitating cross-morphology reuse of manipulation data and policies.
DeFI: Decoupling forward prediction and inverse dynamics so robots can actually learn from large-scale video · vla
Learning policies from unannotated video often suffers from "objective misalignment" — models predict visual frames but don't learn executable actions. DeFI uses a decoupled design to separately handle "predict what will happen" and "decide what action to take."
ICLR 2026 · Commentary: The Embodied Way source (WeChat, CN)
Addressing the objective-misalignment problem that arises when existing VLAs learn directly from video, DeFI decouples forward state prediction from inverse dynamics (inferring actions from state transitions), assigning "understanding how the world changes" and "outputting executable actions" to separate modules — more efficiently exploiting large-scale video data.
COMAP: Using world models to compensate for LLMs' weakness in dynamic reasoning · world-model
Language models excel at static knowledge but struggle with tasks that require simulating consequences. COMAP lets a world model predict future states for candidate actions and then feeds that back to improve decision-making — a concrete path toward "world models empowering agent planning."
Commentary: Jiqizhixin (Chinese AI media) source (WeChat, CN)
The world model predicts future states for candidate actions; the agent uses these predictions to optimize its actions; the resulting trajectories are then fed back via self-distillation to update the world model, forming a closed loop. The paper reports that COMAP achieves approximately 16.75% relative improvement on Qwen3-4B across embodied task planning, web navigation, and tool-use benchmarks.
NTU: First 3D generative model with physical simulation support — generated assets can go directly into robot training · world-model
Generative 3D assets have mostly been limited to "looks good" rather than "usable" in physical interaction. Embedding simulatable physical properties into the generation step means the output can feed directly into robot training, closing the last gap in synthetic data pipelines.
Nanyang Technological University (NTU) · Commentary: DeepTech source (WeChat, CN)
The team's 3D generative model produces geometric assets along with properties usable for physical simulation, enabling generated results to be deployed directly in robot training and simulation environments rather than serving only as visual assets — pointing toward a "generation as simulation data source" workflow.
Other papers today: villa-X (Microsoft, uses a Latent Action Model to compress inter-frame visual changes into latent action tokens, strengthening VLA pretraining with zero-shot transfer to unseen embodiments); Fast-dVLA (ECCV 2026, real-time discrete-diffusion VLA inference acceleration); LabVLA (Zhejiang University, VLA designed for scientific laboratory instruments and transparent liquids); EVO-1 (0.77B lightweight VLA, 2.3 GB VRAM, 16.4 Hz, already deployed on L'Oréal production lines via Qingcang Robotics (Chinese robotics company)).
Open Source · Tools · Benchmarks
· Unitree RL Lab: Unitree's (Chinese robotics company) official open-source reinforcement learning training environment built on IsaacLab, supporting the Go2/H1/G1 platforms with a complete sim-to-real pipeline.
· ACE-Ego: A "one brain, multiple embodiments" manipulation VLA open-sourced jointly by Daxiao Robotics (Chinese robotics startup) and CUHK MMLab, claiming top results on two embodied benchmarks (⚠️ benchmark scope; this model was first reported in mid-June).
II. Funding & Deals
NEURA Robotics ｜ Series C ｜ Up to $1.4 billion ｜ Valuation ~$7 billion · humanoid
Announced earlier this month and widely covered in the Chinese press this week. Investors include Tether, Qualcomm, Amazon, Nvidia, Bosch, Schaeffler, and the European Investment Bank; the full amount is committed subject to performance milestones. The company states its order backlog and deployment pipeline already exceeds $1 billion, with a target of producing millions of units by 2030. This is the largest single funding round disclosed to date by a full-stack robotics company globally, and lifts the capital narrative for European robotics to a new level.Source: NEURA Robotics / CNBC
Bear Robotics acquires Kinisi Robotics ｜ M&A · embodied
Bear Robotics (food service and delivery AMR company, with more than 16,000 units deployed globally) has signed a definitive agreement to acquire UK-based Kinisi Robotics, bringing in its KR1 humanoid robot, the Bristol engineering team, and manipulation AI capabilities. The acquisition fills the "dexterous manipulation" layer beyond Bear's existing "mobility + delivery" stack, completing an end-to-end physical AI platform. Kinisi founder Brennan Pierce will serve as Bear's Chief Robotics Officer, leading ongoing KR1 platform development and manipulation technology integration.Source: ACN Newswire source
Jiangxing Intelligence ｜ Series C & D ｜ Hundreds of millions of RMB (strategic) · industrial
Jiangxing Intelligence (Chinese edge-AI and industrial automation startup), focused on "physical AI at scale," has closed Series C and D strategic rounds totaling hundreds of millions of RMB. The company originated in edge intelligence and device autonomy for industrial sites; these rounds continue a pattern of industrial-capital and strategic-investor participation, reflecting a shift in funding from "building the brain / building the body" toward the application layer of deploying physical AI in real factory settings.Source: Sohu source
LISSOME ｜ Series A ｜ Tens of millions of RMB · embodied
LISSOME (Chinese AI kitchen robot startup), focused on AI-powered kitchen robotics, has closed a Series A round of tens of millions of RMB. Food service and back-of-house operations represent a relatively structured embodied-manipulation segment with clear willingness to pay, and the space has been attracting sustained capital attention.Source: Dahe Cube source
Companion robot startup ｜ Angel round ｜ Tens of millions of RMB · adjacent
A companion-robot team formed by former executives from Moody (Chinese consumer electronics brand) and core staff from DJI has closed an angel round of tens of millions of RMB, led by Jinqiu Fund. Emotional companionship is widely regarded as one of the most commercially viable segments in consumer robotics, and the team's consumer electronics and drone supply-chain background is seen as an advantage for mass production.Source: 36Kr source (WeChat, CN)
III. Commercial Deployment
Pony.ai fully opens Singapore Robotaxi service to the public · autonomy
Pony.ai (Chinese autonomous driving company) has announced that its autonomous ride-hailing service in Singapore is now fully open to the public, with expanded reach through integration into the Zig ride-hailing app. Following deployments across multiple cities in China and expansion into the Middle East, this is a substantive step — moving from "testing / pilot operations" to "open to the general public" in Southeast Asia — distinct from the many recent global expansion announcements that remain at the MoU or agreement stage.Source: Gasgoo source
General Motors deploys ~50 robots at Factory Zero; UAW protests after 1,000+ job cuts · industrial
General Motors has deployed roughly 50 robots at its Factory Zero electric vehicle plant in Detroit, after previously eliminating more than 1,000 positions at the facility. The United Auto Workers (UAW) union has publicly condemned the move, and both sides have clashed sharply over automation-driven job displacement. It should be noted that the "50 robots vs. 1,000+ job cuts" framing is primarily a union and media construct — the two are not in a strict one-to-one substitution relationship. Nonetheless, the episode has put the employment impact of factory automation back in the spotlight and is one of the most prominent social debates to emerge from this wave of manufacturing robotics.Source: NDTV source
Alibaba's logistics unit rolls out rack-climbing warehouse robot · industrial
Alibaba's logistics unit has unveiled a warehouse robot capable of climbing shelving racks to retrieve and place items, targeting the pain points of high-position storage access and space utilization — moving into actual deployment in e-commerce warehouse automation. Specific deployment scale and throughput figures should be verified against official announcements.Source: Benzinga source
China's first industrial embodied inspection robot enters mass production in Luoyang; Zhiyuan Robotics' Elf G2 begins 6-day production-line inspection livestream · embodied
An industrial embodied inspection robot developed in Luoyang, China has announced mass production, targeting the high-frequency, clearly defined need for visual and defect inspection in manufacturing. Concurrently, Zhiyuan Robotics (Chinese humanoid robot company) announced a June 23–28 joint livestream with Longqi Technology, covering the Elf G2 humanoid robot across 8 tablet-inspection workstations on a production line. ⚠️ Deployment/livestream claims Production scale and the degree to which this constitutes "routine operations" still need to be verified by subsequent installation and utilization data; the livestream itself serves both demonstration and marketing purposes.Source: Robot Insights source (WeChat, CN)
UBTECH's "YouWorld" ultra-bionic humanoid accumulates 5,000+ pre-orders on JD.com · humanoid
UBTECH's (Chinese humanoid robot company) "YouWorld" ultra-bionic humanoid robot launched for pre-order on JD.com, with disclosed cumulative pre-orders exceeding 5,000 units. ⚠️ Pre-order figures Pre-orders (typically refundable deposits) reflect market enthusiasm rather than actual deliveries; final conversion and fulfillment figures will only be confirmed once units ship.Source: Jrj.com source
IV. Industry Developments
Nvidia launches Halos for Robotics: Industry's first full-stack physical AI safety system · adjacent
Nvidia has extended its Halos safety architecture — originally developed for autonomous driving — to robotics, introducing what it claims is the "industry's first full-stack, open" physical AI safety system, providing a unified safety framework for machines that can perceive, decide, and act. The system has three layers: IGX Thor + Holoscan Sensor Bridge for AI compute and sensor ingestion; the Halos OS software stack for safety functions and applications; and the Halos AI Systems Inspection Lab to help partners connect with third-party certification bodies. Agility Robotics is the first humanoid and physical AI company to integrate its proprietary safety system; Hesai Technology (Chinese lidar company) has joined the inspection lab; and FORT Robotics is extending "Outside-In" safety capabilities on top of Halos. As humanoid robots accelerate into factories and warehouses, safety and certification are shifting from an optional extra to a foundational prerequisite for scaled deployment — and the race to define that standard layer is becoming a new strategic battleground.Source: GlobeNewswire source
Haiqing Zhiyuan lists on Hong Kong Stock Exchange, dubbed "first physical AI IPO"; shares surge 300%+ at open · adjacent
Shenzhen-based Haiqing Zhiyuan (01392.HK; Chinese multispectral sensing and AI company) listed on the main board of the Hong Kong Stock Exchange, surging more than 300% at the open with an intraday market capitalization of approximately HK$22.5 billion, earning the market label "first physical AI IPO / first multispectral AI IPO." To be clear, the company's core business is multispectral (infrared/ultraviolet) sensing + AI algorithm fusion — closer to the perception layer than to complete embodied robots or manipulation brains. The "physical AI" label is largely a market narrative classification. That said, as the first high-profile listing under the physical AI banner, its debut performance reflects the intense investor enthusiasm for the sector.Source: National Business Daily source
UBTECH launches Walker C1 commercial service humanoid at China International Supply Chain Expo · humanoid
UBTECH has unveiled the Walker C1, a new-generation humanoid robot for commercial service applications spanning reception and wayfinding, commercial services, entertainment, and educational research in urban settings. The robot appeared as the expo's "first silicon-based ambassador," with an on-site human-robot dance demonstration. ⚠️ Vendor launch/demo The real-world value of commercial service humanoids ultimately depends on reliable sustained operation and unit economics; the launch and dance demonstration are capability showcases, and deployment quality will require field data.Source: Securities Times source
Robotaxi goes right-hand drive: WeRide × Geely Yuancheng × Kwoon Chung Bus sign agreement in Hong Kong · autonomy
Following recent expansion into multiple European cities, WeRide (Chinese autonomous driving company) has signed an agreement with Geely Yuancheng and Kwoon Chung Bus at the Hong Kong Auto Show to jointly develop mass-production Robotaxi vehicles for right-hand-drive markets, targeting Singapore, the UK, Japan, and Australia. ⚠️ Partnership/plan announcement This is a joint development agreement; production vehicle specifications and rollout timelines have not yet been announced. This is an extension of the Robotaxi globalization storyline rather than a new operational milestone.Source: WeRide source (WeChat, CN)
Japan plans to mobilize ~$65 billion in "physical AI" investment by 2040 · adjacent
According to the Nikkei, Japan plans to channel approximately ¥10.5 trillion (~$65 billion) into "physical AI" through public-private partnerships across 17 strategic sectors by fiscal year 2040, with a stated goal of capturing more than 30% of the global AI robotics market by 2040. The announcement lifted Tokyo technology stocks and pushed the Nikkei temporarily above 72,000. ⚠️ Government planning This represents a long-horizon investment framework and target; implementation and fund allocation are subject to future policy detail.Source: Nikkei Asia
Automate 2026: A flurry of industrial robot platform partnerships announced · industrial
Automate 2026, North America's largest robotics trade show, served as a hub for new industrial automation products and alliances: Vention announced separate partnerships with FANUC America and Teradyne Robotics (UR), connecting industrial robots and digital twins to its AI hardware-software platform to accelerate Universal Robots cell deployment; Alphabet's Intrinsic showcased next-generation modular industrial AI assembly; Doosan Robotics unveiled an AI palletizing solution; Cobot's Proxie Gen 2 adds autonomous task orchestration and mobile manipulation. The throughline is "making industrial robots easier to orchestrate and faster to deploy," with platform integration and plug-and-play becoming key competitive battlegrounds.Source: The Robot Report source
Ruqi Mobility (Chinese ride-hailing and AV data company) launches embodied AI data platform, continuing "data as the differentiator" strategy · adjacent
Ruqi Mobility has launched an embodied AI data platform emphasizing "out-of-the-box usability," aiming to reduce the marginal cost of first-person data from collection to training and expand AI data service boundaries; Yuanqi Innovation has taken a strategic stake as a co-builder. Multiple outlets are covering the story in tandem, reflecting a broader industry trend of positioning "data infrastructure" as a standalone segment alongside hardware and AI models.Source: East Money source
World model PAIWorld claims top ranking on WorldArena leaderboard · world-model
The physical intelligence team at the Chinese Academy of Sciences Institute of Industrial Artificial Intelligence claims its in-house world model PAIWorld has reached the top of the international WorldArena leaderboard in a recent update. ⚠️ Leaderboard claim This is a self-reported result; the authority and evaluation methodology of the WorldArena benchmark itself remain debated. Treat this as a signal of intensifying competition in world models rather than a definitive capability ranking.Source: Jiqizhixin source (WeChat, CN)
Hardware · Supply Chain
· Dexterous hand drive modules (Weihong Co. / Hamm Electronics): The dexterous hand drive modules from the acquired Hamm Electronics (Chinese motion control component maker) have progressed from sampling to small-batch and mass production, with at least 600 motors or modules shipped daily since May; customers include Lingxin Qiaoshou, Yinshi, Qiangnao, and Zibianliang (all Chinese robotics or neurotech companies) (⚠️ company-reported figures).
· Dexterous hand cost structure: Dexterous hands account for roughly 17% of a humanoid robot's BOM and are among the most expensive components; the industry has raised approximately RMB 8.7 billion in aggregate funding yet no company has achieved profitability, caught in the "performance–cost–reliability" trilemma.
· Chinese supply chain cost reduction: Reports indicate Chinese suppliers have brought dexterous hand prices down from approximately RMB 2 million to approximately RMB 50,000 (⚠️ reported figures; reflects variation across DOF levels and configurations).
· Component supply: Shangluo Electronics (Chinese electronic components distributor) states it has begun supplying passive components such as resistors and capacitors to Unitree Robotics.
