Today's Highlights
· After 4 months of on-the-job training in a car factory, Xiaomi's humanoid robot's nut-fastening success rate has risen to 98%, just one percentage point short of the human benchmark
· Amap released and open-sourced the general-purpose world model workshop ABot-World Studio, which can run locally on a single RTX 5090 with continuous inference for over an hour
· The HiPhi (Huawei-backed EV brand) G9 became the first vehicle in Beijing to receive an L3 autonomous driving road test license rated for 120 km/h
· Two tactile/dexterous-hand deals in one day: Xense Robotics raised over 100 million yuan, and DexRobot's Series A of several hundred million yuan was led by Shanghai Electric
· Xiaomi's U0 treats the world model as a data factory, lifting pi0.5's out-of-distribution success rate from 36.9% to 63.2%
I. Papers
Xiaomi-Robotics-U0: World models don't belong in the control loop — use them as a robot data factory instead · world-model
This paper offers a pragmatic answer to "how should world models actually be used in robotics": rather than doing real-time planning for the robot at inference time, it "manufactures" scenarios the robot has never seen before training. Xiaomi (Chinese consumer electronics and EV company) proposes a unified embodied synthesis framework that uses a world foundation model to generate multi-view-consistent, geometrically coherent, embodiment-constrained interaction data — directly filling the most expensive gap in real-robot data: out-of-distribution scenarios. The result: pi0.5's out-of-distribution task success rate rose from 36.9% to 63.2%, nearly doubling, with no additional real-robot data collection.
Xinghang Li et al. (Xiaomi) · arXiv 2607.11643 source · Commentary: ICCV source (WeChat, CN)
ABot-N1: Turning vision-language navigation into a general-purpose foundation model · autonomy
The most-discussed paper of the day in the community (HF↑76). Most existing navigation foundation models are monolithic policies that map observations directly to actions, which are prone to coordinate drift and struggle with long-tail semantics. ABot-N1 separates deep spatial reasoning from multi-task generality and then unifies them, aiming for a VLN backbone that generalizes across embodiments and scenarios. It belongs to the same ABot series that Amap (Alibaba's mapping subsidiary) released together today, alongside the two papers below.
Ruiyan Gong et al. (Amap) · arXiv 2607.10383 source
ABot-3DWorld 0: Turning a sentence, image, or video into an explorable 3D world · world-model
This is the 3D generation engine behind Amap's ABot-World Studio, which launched today — the model has been open-sourced. Its core is a compact representation called the Spatial Generative Primitive (SGP) — a high-quality panorama paired with a spatial point cloud — used to uniformly describe arbitrary 3D spaces. This converges text, image, and video inputs into a single, real-time-explorable world representation, directly serving robot training and simulation data collection.
Mingchao Sun et al. (Amap) · arXiv 2607.11673 source
ABot-AgentOS: Giving robots an "operating system" with lifelong memory · benchmark
HF↑66. VLM/VLA models have advanced perception and action prediction, but long-horizon tasks still lack a general runtime layer — someone still has to manage reasoning, memory, tool calling, result verification, and cross-embodiment execution. This paper abstracts that layer into a general-purpose robot Agent OS with cross-modal lifelong memory — addressing a gap in "upper-layer orchestration" that papers tend to overlook but that shows up daily in real-world deployment.
Jiayi Tian et al. (Amap) · arXiv 2607.10350 source
EgoSteer: Feeding dexterous hands "steerability" from first-person human video · vla
Steerability — doing what you're told, and changing course when told to — is a capability general-purpose policies should have but dexterous-hand systems largely lack. The bottleneck is data: there's a shortage of large-scale, language-aligned, action-precise demonstrations. EgoSteer offers a full-stack solution that uses first-person human video to scale up pretraining for dexterous-hand VLA models, followed by post-training with a very small amount of real-robot data.·
Yifan Zhong et al. · arXiv 2607.09701 source
Mixture of Frames Policy: Action denoising shouldn't happen in just one coordinate frame · manipulation
Manipulation is inherently multi-frame: fine end-effector motion is simplest to express in the tool frame, while transport, righting, and whole-body coordination make more sense in the base frame. This paper has the policy denoise separately in multiple reference frames and then blend the results, targeting bimanual mobile manipulation — a structural correction to diffusion/flow policy action representations rather than just more parameters.
Dian Wang et al. (Stanford) · arXiv 2607.11884 source
VIA: Letting foundation models control robots by "looking at" and "clicking on" an interface · vla
General foundation models already have strong vision and reasoning; their weak point is closed-loop control. Sadigh's group's approach isn't to hard-train an end-to-end policy, but to give the foundation model a visual interaction interface as a middle layer, letting it operate the robot agent-style — realigning "what the model is good at" with "what the robot needs."
Hengyuan Hu et al. (Stanford) · arXiv 2607.11119 source
Cycle-World: Using "backward prediction" to curb drift in long-horizon video world models · world-model
In autoregressive diffusion for long video generation, errors accumulate step by step into structural collapse — a wall every team hoping to use video world models as simulators runs into. This paper adds a backward-prediction cycle-consistency constraint, letting the model check for itself whether "the present can be recovered from the future," thereby suppressing long-horizon generation drift.
Zihan Su et al. · arXiv 2607.11836 source
Other papers today: Towards Human-level Dexterous Teleoperation, a teleoperation system for in-hand regrasping and finger gaiting (arXiv 2607.11481 source); WALA, learning executable latent actions simultaneously from action-labeled demonstrations and action-free video (arXiv 2607.11397 source); WarpMPC, solving a hundred thousand MPC subproblems in parallel on a GPU (arXiv 2607.11603 source); TAC-LOCO, integrating touch into quadruped whole-body control for dynamic loco-manipulation (arXiv 2607.10132 source); A Minimalist Retargeting-Guided RL Recipe, testing whether the "retargeting + RL" recipe from humanoid whole-body control transfers to dexterous manipulation (arXiv 2607.11874 source); From World Action Models to Embodied Brains, a roadmap from world-action models toward open-world physical intelligence (arXiv 2607.11689 source).
Open Source · Tools · Benchmarks
· Horizon Robotics simulation-training generation engine: open-sourced; the same scene can run across MuJoCo, Genesis, and IsaacSim simultaneously — the first time robot training can reuse a single scene asset across simulators source (WeChat, CN)
· ABot-World series: alongside ABot-World Studio, Amap open-sourced two model weights — ABot-World0 (interactive video generation) and ABot-3DWorld0 (3D scene generation) source (WeChat, CN)
· GigaWorld-1: a cost-reduction approach for real-robot policy evaluation, using generative world models to replace part of real-robot trial runs source (WeChat, CN)
II. Funding & Deals
Xense Robotics (千觉机器人) | New round | Over 100 million yuan · hardware
Strategic investment from a top embodied-intelligence industry player and Chigo (Chinese home appliance maker), with follow-on from Tianji Capital. Founded in May 2024 by Ma Daolin, an associate professor at Shanghai Jiao Tong University and the sole global winner of the ICRA 2021 Best Paper Award, the company builds an integrated solution combining multimodal tactile sensors with vision-tactile data and models, and says it has served over 300 leading industry clients. The funding will go toward industrializing high-precision, three-color-light vision-tactile sensors. Touch sensing is moving from "a modality in papers" to a component with real orders and a place in the supply chain — this is the second tactile-sensing company to raise over 100 million yuan this week.Source: Sensor Expert Network source (WeChat, CN)
DexRobot (灵巧智能) | Series A | Several hundred million yuan · embodied
Investment from Shanghai Electric. The company specializes in robot end-effectors and dexterous manipulation systems, describing its mission as solving the "last centimeter" of a robot's interaction with the physical world. The funding will go toward product iteration for its dexterous hands and scaled deployment across multiple industries. Industrial capital (rather than pure financial VC) betting directly on end-effectors points to real industrial demand for "hands that can do fine work."Source: Robotics Open Community source (WeChat, CN)
LimX Dynamics (逐际动力) | Pre-IPO round | Nearly $200 million | Post-money valuation 15 billion yuan · humanoid
Following disclosure of this round earlier in the week, the company officially confirmed closing and finalized the valuation today. Investors include IDG Capital, Lens Technology, Hefei Binhu Financial Holdings Group, GGG Group and Redstone VC, Huashan Capital, and Legend Capital, among others; the company says it has raised $400 million cumulatively over the past six months. For context: Crunchbase data shows global robotics startup funding reached $18.8 billion in the first half of 2026, already surpassing the full-year 2025 total of $15 billion.Source: Shanghai Securities News source (WeChat, CN)
Quadric | Series C | $46 million · hardware
Builds a programmable AI chip platform for edge inference. Robot compute solutions still rely heavily on general-purpose GPUs, and programmable edge chips are one path around that dependency.Source: Pulse 2.0 source
NovaTech (South Korea) | New round | 7 billion won (~$5 million) · adjacent
With participation from Hyundai Motor Group. The company builds robot orchestration software, coordinating multiple robot brands via its PiPER platform; this round's funding will go toward expanding into the North American market. An automaker betting on the "orchestration layer" rather than the robot itself is a side signal that multi-robot coordination on production lines is a real pain point.Source: Chosunbiz source
Ondas acquires DZYNE | $875.8 million · adjacent
Unmanned systems company Ondas acquired defense drone maker DZYNE, expanding its defense capabilities. This is the largest robotics-related acquisition of the week by value.Source: The Robot Report source
III. Commercialization & Deployment
Xiaomi's humanoid robot completes 4 months of "internship" in a car factory, nut-fastening success rate hits 98% · humanoid
Lei Jun (Xiaomi founder) disclosed the latest data on the humanoid robot's on-the-job training at the factory: at the self-tapping nut installation station, the bilateral operation success rate rose from 90.2% to 98% over 4 months of iteration — just one percentage point short of the human-operator pass rate. More notably, newly unlocked stations — installing the center console side panel and folding/recycling parts bins — both reached 90% success rates; the side panel is large, irregularly shaped, and highly flexible, with the task involving multiple pick-and-place cycles at multiple positions, marking the first time a robot has completed long-duration handling of a flexible workpiece on a real production line. What makes this significant isn't the 98% figure itself, but that it's one of the rare publicly disclosed, time-series production-line training curves: over the past year, the industry has mostly offered single demos or "X hours without human intervention," whereas "same station, four months, success rate climbing from 90.2 to 98" provides a previously almost never-disclosed metric — the rate of iteration. Of course, 98% is still a single-station training result, still some distance from production-line metrics like overall cycle time, uptime, and defect cost.Source: QbitAI (Chinese tech media) source (WeChat, CN)
DEEP Robotics' quadruped robot deployed at Switzerland's Leibstadt nuclear power plant · industrial
DEEP Robotics (云深处) announced its quadruped robot has been deployed at Switzerland's Leibstadt nuclear power plant, handling digitized inspection and maintenance — replacing human workers on instrument readings, thermal imaging, and vibration baseline collection in radiation zones. Nuclear power is one of the few scenarios where quadruped robots have already proven a viable paying-customer logic (humans shouldn't linger, paths are fixed, data is structurable), and it's a relatively realistic entry point for Chinese robot makers into European industrial clients.Source: ACCESS Newswire source
Tesla expands Robotaxi to Miami, develops wheelchair-accessible model in Texas · autonomy
Tesla's Robotaxi service launched in Miami this week, and the company confirmed it is developing a wheelchair-accessible autonomous vehicle model in Texas. The accessible model isn't just a product-line addition: many U.S. jurisdictions require ride-hailing operators to maintain a minimum proportion of accessible vehicles in their fleets, and Robotaxi can't truly replace taxi fleets without clearing this bar.Source: Electrek source
Turkey's Tripy Mobility to deploy Robotaxis using WeRide-Vision (Chinese autonomous driving supplier) tech · autonomy
Chinese autonomous driving solutions provider WeRide-Vision (天瞳威视) has won the Robotaxi project for Turkish mobility platform Tripy Mobility. Chinese autonomous driving solutions are moving beyond exporting vehicles or licensing algorithms toward the deeper commitment of supplying full-stack solutions to local operators.Source: Sina Finance source
Tesco trials 5-foot shelf-inspection robot · industrial
UK retail giant Tesco has begun trialing a 5-foot-tall aisle-inspection robot that automatically identifies out-of-stock shelf locations. Retail inspection is one of the few mobile-robot scenarios that can create value through perception alone, without grasping — but it remains at the trial stage, with no large-scale procurement yet.Source: The Grocer source
Serbia to launch China-backed humanoid robot production in August · humanoid
Serbia will launch a humanoid robot production line with Chinese backing in August. ⚠️ Plan stage Production scale and the specific robot model have not yet been disclosed; the outcome remains to be verified.Source: IntelliNews source
IV. Industry News
Amap releases general-purpose world model workshop ABot-World Studio, all models open-sourced · world-model
Alibaba's Amap officially launched ABot-World Studio today and opened it for testing, while also open-sourcing the two models behind it: ABot-World0 (interactive video generation) and ABot-3DWorld0 (3D scene generation). The product's technical framing is clear: world models have previously split into two separate paths — interactive video generation offers strong immersion but interaction ends once the video finishes, while 3DGS scene generation offers high fidelity but isolated, non-interconnected worlds. ABot-World Studio brings both paths into a single platform: users input text or an image and get a real-time-interactive, shareable world, which can be saved either as video or as a 3DGS file. It also includes a "portal" mechanism — passing through a door jumps you into another complete 3D world, weaving previously isolated scenes into an infinitely explorable network. Two engineering figures are worth industry attention: the company says it can be deployed locally on a single RTX 5090, and a single continuous inference session runs stably for over 1 hour without crashing or quality degradation — whereas comparable world models today typically generate for around 1 minute. For embodied-AI teams, this means the barrier to simulation-based data collection has dropped from a GPU cluster to a single card.Source: Amap source (WeChat, CN)
HiPhi G9 approved for Beijing L3 road test license, first model rated for 120 km/h · autonomy
The HarmonyOS Intelligent Mobility HiPhi G9 has officially received a Beijing municipal L3 autonomous driving road test license, becoming the first vehicle model in Beijing to qualify for L3 testing at 120 km/h. Its technical approach uses multi-sensor fusion combining LiDAR, vision, and millimeter-wave radar, with a Huawei 896-line LiDAR unit standard across the lineup; it has passed simulation testing, closed-course testing, and four rounds of validation covering functional safety, safety of the intended function, cybersecurity, and data security. Its L3 operational design domain covers nighttime scenarios and supports lane changes on highways/expressways. Viewed against last week's publicized mandatory national standard for autonomous driving (effective 2027, requiring heterogeneous redundant sensing for L3), the fusion-sensing approach's first-mover advantage in China's L3 approval process is now being converted into an actual license.Source: Yicai source
ByteDance's head of generative AI, Yuan Zehuan, departs to start a world-model venture · world-model
Yuan Zehuan, who previously led ByteDance's generative AI direction, has left to start a venture in the world-model space. ⚠️ Single source Combined with this week's report that ByteDance's Seed world-model team is exploring autonomous driving (officially denied as having any business plan), world models are simultaneously drawing away both capital and talent from big tech companies.Source: AIcore (Chinese tech media) source (WeChat, CN)
Beijing builds and opens embodied-AI robot pilot validation platform for the power industry · industrial
An embodied-AI robot pilot validation platform for the power industry has been built in Beijing and is now officially open for external service. Pilot platforms address what robot makers most lack — batch validation under real power-grid conditions (high voltage, live wires, narrow corridors) rather than benchmark testing in a lab. Such industry pilot facilities, led by grid operators and power research institutes, are becoming the actual chokepoint through which embodied AI must pass on its way from demo to industry admission.Source: Guandian.cn source
JD's robotics ecosystem adds two more partnerships: Huilun Technology, Luming Robotics · adjacent
Following the groundbreaking of its Guangzhou RoboBase, JD (Chinese e-commerce giant) signed strategic partnerships this week with embodied-AI companies Huilun Technology and Luming Robotics; Luming disclosed a partnership target of 1 billion yuan in sales over 3 years, with multiple robot hardware products set to launch on JD in 2026. JD's approach differs from both robot makers and algorithm companies: trading channel and supply-chain access for ecosystem position, positioning itself as the sales and infrastructure backbone for robotics.Source: Gasgoo source
14 institutions across Japan, the US, and Europe join forces on physical AI research; Bengio joins Noetra · adjacent
Led by Japan, 14 institutions across the US, Europe, and Japan have formed a physical AI research collaboration network aimed at building a Japan-native physical AI model; Yoshua Bengio has joined the affiliated company Noetra to help drive its commercialization. Japan's underlying motivation here is easy to read: its robot hardware and components remain world-class, but its "brain" layer — model capability — now clearly lags behind China and the US.Source: The Japan News source
South Korea's SKT and KT launch AI-RAN trial networks to lay groundwork for physical AI services · adjacent
South Korea's two largest carriers, SK Telecom and KT, are deploying "AI base stations" at industrial sites, launching an AI-RAN pilot worth roughly $11.5 million as a core piece of South Korea's "AI highway" initiative. Carriers are treating physical AI as the next growth driver after 5G/6G — robots on factory floors need to upload video and receive policy updates in real time, and the network layer is indeed an often-overlooked bottleneck.Source: Seoul Economic Daily source
Hardware · Supply Chain
· Dedicated dexterous-hand chips: CAS Wireless Semiconductor (中科无线半导体) has launched China's first integrated dexterous-hand chip solution, the CT-2001H/CT-1906H, paired with a CT-HS01 4D spectral sensor achieving 0.1mm positioning precision and microsecond-level latency; the company says it cuts hardware costs by 30% and is natively compatible with VLA models source (WeChat, CN)
· Boya Intelligence (伯牙智能) Gaoshan S1: a dexterous hand weighing just 500 grams with 30N of fingertip force, developed by a team co-founded by a former Alibaba executive and a robotics professor source (WeChat, CN)
· Yuequan Bionics (月泉仿生) Yingshou Y-Hand M1: the full hand weighs 299.7 grams; rather than competing on degrees-of-freedom specs, it uses a bionic tension-compression transmission design source (WeChat, CN)
Top comments (0)