Today's Highlights
· Ant LingBot (Ant Group's embodied AI team) closes today's roster with LingBot-VA 2.0, formally proposing "embodied-native" rearchitecture for robot brains
· 1X unveils a 25-DOF tendon-driven dexterous hand, NEO's Hands, claiming near-human strength and dexterity
· Zhiyuan Robotics (Chinese humanoid startup) says its C5 cleaning robots have logged 70 units running at BYD factories, with annual savings near 10 million yuan per unit claimed
· Tactile sensor maker Tashan Technology (Chinese tactile-sensor startup) closes a Series B worth several hundred million yuan, with in-hand orders in H1 already 4x full-year 2025
· China's Ministry of Industry and Information Technology and SASAC launch a real-world training program for humanoid robots, targeting 10,000+ unit deployment by year-end
I. Research Progress
LingBot-VA 2.0: A Video-Action Foundation Model Built for Embodiment · world-model
This is the culmination of Ant LingBot's "Embodied Native" paradigm and today's headline paper — it puts a stake in the ground on a dividing question for the field: should robot brains follow the old internet-scale-model playbook of "transfer plus fine-tune," or be rebuilt from scratch around the physical world? The authors argue for the latter: video generation models are designed for digital content and are inherently ill-suited to physical environments, so they rebuilt everything from the data and training objectives to the architecture. Methodologically, a semantic visual-action tokenizer aligns visual representations with actions, paired with causal pretraining, an MoE architecture, and "look-ahead reasoning" that merges prediction with execution and compresses inference cost below the real-time control threshold. The model is released open-source alongside a technical report, Apache-2.0 code, and a 6B-parameter checkpoint, emphasizing instruction-following, action precision, and cross-embodiment generalization.
Qihang Zhang et al. (Ant LingBot / Robbyant) · arXiv 2607.08639 source · via Embodied AI Frontier source (WeChat, CN)
EgoWAM: Training World-Action Models on "In-the-Wild" First-Person Human Video · world-model
First-person human video is a cheap, scalable source of supervision for robot manipulation, but behavior cloning tends to entangle transferable object, scene, and task semantics with non-transferable human body morphology, head motion, and behavioral style. EgoWAM holds the policy backbone, action head, and data mix fixed while varying only the "world prediction objective," systematically comparing pixel-, DINO-, and 3D-motion-flow-based targets to clarify which world-prediction objective best aids transfer in human-robot co-training.
Baoyu Li et al. (including members of Danfei Xu's group) · arXiv 2607.08436 source
Structural Flaws in Language-Conditioned World Models, and a "Write-Protection" Fix · world-model
Directly targets the mainstream approach — used by RT-2, Octo, and PaLM-E — of feeding LLM/VLM features end-to-end into robot world models: the paper argues that letting language gradients directly shape physical symbolic representations causes "symbol collapse," and proposes a sufficient fix via a "write-protected discrete bottleneck" — achieving zero symbol collapse across 32 random seeds and 79–100% semantic grounding across three encoders (CNN, V-JEPA 300M, CLIP ViT-L) and two environments. A rare "cold-water-plus-fix" mechanistic paper among today's world-model entries.
Jiayi Fang · arXiv 2607.08312 source
FabriVLA: A 1B-Parameter Small Model That Handles Multi-Task Manipulation · benchmark
Attaches a flow-matching action head to an InternVL3.5 vision-language backbone, achieving a 90.0% hierarchical average success rate on Meta-World MT50, which spans 50 manipulation tasks — showing that a compact VLA built on a 1B-class VLM can perform strongly without relying on a multi-billion-parameter backbone, favorable for inference latency in real-robot deployment.
Shiyuan Yang et al. · arXiv 2607.08575 source
ContactMimic: Teaching Humanoids to Learn Object Interaction via "Contact Commands" · manipulation
For tasks like sitting in a chair, wiping a whiteboard, or pushing furniture, tracking keypoints alone isn't enough — a robot could strike the right pose without ever making meaningful contact with the object. Beyond keypoint trajectories, ContactMimic additionally tracks part-level binary contact commands, using a contact-following reward and trajectory augmentation to break the spurious correlation between keypoints and contact, so humanoids can genuinely press against and push objects.
Xinyao Li et al. (UIUC, Saurabh Gupta's group) · arXiv 2607.08742 source
Latent Memory Palace: Bringing "Think-on-Demand" to Continuous Control · vla
Human decision-making is sometimes instantaneous and sometimes deliberative, and language models have shown similar adaptive "reasoning" — but transferring this to continuous control policies has proven difficult, since reasoning directly in language space lacks spatial granularity. This work shows that organizing information into an autoregressive latent space (resembling a "memory palace"), with adaptive iterative retrieval, lets control policies "emerge" with reasoning capability as well.
Chuning Zhu et al. (University of Washington / Toyota Research Institute) · arXiv 2607.08724 source
Feasibility Study on Humanoid Robot Live Surgery (Unitree G1, previously reported) · manipulation
The technical paper behind what was already yesterday's headline — the world's first live surgery performed by a humanoid robot: a general-purpose humanoid fitted with commercial wristed laparoscopic instruments for teleoperation, validated through bench characterization, dry-lab multi-level user studies, and live porcine trials. Multiple outlets covered the same story again today; as a previously reported milestone, this entry serves only to note the paper's provenance.
Zekai Liang et al. (UC San Diego) · arXiv 2607.07972 source · via Caijing.com source (WeChat, CN)
Other papers today: Understanding and Mitigating the Video-Action Generalization Gap — inference-time adaptive guidance to mitigate compositional generalization degradation after fine-tuning video-action models (arXiv 2607.08127 source); AnyDexRT, calibration-free dexterous-hand teleoperation retargeting (arXiv 2607.08341 source); DexVerse, a modular benchmark for multi-task, multi-embodiment dexterous manipulation (arXiv 2607.08751 source); WCog-VLA, a dual-layer "world-cognition" VLA for end-to-end autonomous driving (arXiv 2607.08375 source); Post-Training in End-to-End Autonomous Driving, a survey (arXiv 2607.08072 source); physics-guided gait adaptation for humanoids on steep slopes (arXiv 2607.07830 source); GIRAF, generalizable full-body human-articulated-object interaction generation from a Meta team (arXiv 2607.07880 source).
Open Source · Tools · Benchmarks
· RLinf: open-source reinforcement learning framework adds RLT algorithm support, emphasizing synchronized sim-to-real verifiability source (WeChat, CN)
· Robo-ValueRL: open-source value-alignment RL framework for humanoid robots released, moving from "not knowing better" to "telling right from wrong" source
· dVLA-RL: released jointly by Shanghai Jiao Tong University and Baidu's Baige (AI computing platform), opening a path toward autonomous evolution for robot foundation models source
· Cross-embodiment transfer infrastructure: a CMU team builds missing infrastructure for "moving AI across different robots" source
· UNIBOT World Challenge: Unitree launches a Physical AI challenge, evaluating generalization with 32 tabletop manipulation tasks source
II. Funding & Deals
Tashan Technology | Series B | several hundred million yuan · hardware
Pursuing a tactile-sensor route aimed squarely at "tactile-native models." Investors include Taiping Innovation, Joyson Electronics, AUX, and Pengling Co. among industrial and specialized capital, with existing backers Hongshan Capital and Binfu Capital following on. Founded by Tsinghua and Beihang alumni, the company says its tactile sensors now ship at tens of thousands of units per month, with in-hand orders for H1 surging to 4x full-year 2025 — a rare signal of simultaneous capacity and order-book traction in China's push for tactile-sensor self-sufficiency, as touch becomes a holy-grail component for dexterous hands.Source: Robot Lecture Hall source (WeChat, CN)
Lisin Technology | Series B | nearly 500 million yuan · hardware
Racing toward on-device large-model chips, with its Nebula series slated for mass production by year-end. On-device compute is a key bottleneck for cost reduction and real-time inference in embodied AI/robot platforms, and this nearly 500-million-yuan round is a bet on that chain.Source: Sohu source
West Lake Robotics | New round | over 100 million yuan (three rounds totaling several hundred million yuan in 5 months) · embodied
Funded exclusively by Henan Investment Group's Huirong Fund; founder Wang Donglin is deputy director of Westlake University's AI department. The joint backing of a state industrial fund and market-driven institutions, plus three rounds in five months, shows academic-lineage embodied AI teams remain a priority target for capital.Source: Sohu source
Xspark AI | Angel round | nearly 100 million yuan · embodied
Co-led by CDH VGC and Vertex Capital China (formerly Chuxin Capital), pursuing a "trustworthy embodied AI architecture" aimed at making robots reliably operational and accelerating Physical AI at scale.Source: Sina Mobile source
Proception | Seed round | $11 million · embodied
Builds dexterous hands for embodied fine manipulation — 22 degrees of freedom (18 actuated), humanlike tendon-driven design, four joints per finger, emphasizing a "closed-loop evolution for fine manipulation." This overseas dexterous-hand startup lands in the same news cycle as today's 1X hardware buzz.Source: Alpha Commune source (WeChat, CN)
Other funding today: Wujie Zhihang, an angel round worth hundreds of millions of yuan, claims it has already partnered with companies including Nvidia (Source: Dahe Caili Fang source); MScape, a seed round of over 100 million yuan (Source: Gasgoo source).
III. Commercialization & Deployment
Zhiyuan Robotics Declares "Year One of Deployment": C5 Cleaning Robots Log 70 Units at BYD · industrial
Zhiyuan is shifting its narrative from "showmanship" to "the ledger": according to data disclosed at a partner conference and reported by Xinhuanet, BYD Group has deployed roughly 70 units of the Jiechen C5 cleaning robot since April 2025, with each unit replacing more than one cleaning worker and claimed annual savings of nearly 10 million yuan per unit; separately, LY Manufacturing has deployed over 100 units with a payback period under 12 months. Official specs for the C5 list a cleaning efficiency of 1,980 m²/h, a 90L water tank, and a self-cleaning wastewater tank cycle of 4 minutes. The accompanying "mothership data": Zhiyuan's 2025 revenue was 1.05 billion yuan, with a 2026 target of 4 billion yuan; cumulative shipments of its general-purpose embodied robots have grown from 5,000 units by end of 2025 to over 10,000 units by end of March. President Peng Zhihui (known as "Zhihui Jun") declared "2026 is Year One of embodied AI deployment," targeting 1,000 factory deployments this year and 10,000 in industrial settings next year.⚠️ Some figures are company-reported (revenue target and annual savings estimates are company disclosures)Source: Jiazi Zhiqi source (WeChat, CN)
Deep Robotics Quadruped Deployed at a European Nuclear Power Plant · industrial
Deep Robotics' (Chinese quadruped robot maker) four-legged robot has entered a European nuclear power plant for digital operations-and-maintenance inspection — an overseas deployment of a Chinese quadruped robot in a high-hazard industrial setting, pointing toward reduced-staff or unmanned nuclear plant O&M.Source: Sohu source
Zhijian Power Puts 100 Robots on Factory Floors · industrial
In contrast to the "still folding laundry" demo track, Zhijian Power says it already has 100 units in actual production-line work, emphasizing scaled deployment in manufacturing settings.⚠️ Single-source claimSource: Sina Mobile source
Dobot Enters Auto Production Lines with "Industrial AI" · industrial
Dobot is applying industrial AI to automotive production lines to address efficiency pressure in manufacturing — a deployment attempt for robotic arms/collaborative robots in flexible automaker floor work.Source: Sohu source
XPeng GX Robotaxi Enters Pre-Installed Mass Production (previously reported) · autonomy
Following its first internal employee test ride reported yesterday, XPeng's GX pure-vision Robotaxi has entered pre-installed mass production, equipped with 4 self-developed Turing chips and 3,000 TOPS of onboard compute, with the company claiming just 8 months from announcement to mass-production passenger service.⚠️ Company claimSource: Netauto source
IV. Industry Developments
1X Unveils 25-DOF Tendon-Driven Dexterous Hand, NEO's Hands · hardware
OpenAI-backed 1X Technologies has released a next-generation dexterous hand for its NEO humanoid: 25 degrees of freedom (22 fully actuated in fingers and palm, plus 3 at the wrist), tendon-driven with full-DOF back-drivability (pressing a finger yields compliantly and provides force feedback), integrated tactile sensing, IP68-rated dust/water resistance, and compliance with food-safety standards. Peak torque figures: 3.5 N·m at the thumb CMC joint, 2.6 N·m at finger MCP joints, and 17.75 N·m at the wrist. CEO Bernt Bornich claims its dexterity, strength, speed, and reliability approach or exceed human levels, and says the company is preparing to produce 10,000 units. The demo video has drawn millions of views.⚠️ Company claim ("approaching/exceeding human" and the 10,000-unit production plan are company statements, still pending third-party and real-production-line verification)Source: Robot Frontier source (WeChat, CN)
China's MIIT and SASAC Launch 2026 Real-World Humanoid Robot Training Program · humanoid
China's Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission (SASAC) jointly launched the 2026 Real-World Training Program for Humanoid Robots and Embodied AI, setting a target of routine deployment of key humanoid robot products by the end of 2026, distilling 100 high-value application scenarios, and reaching a deployment capacity of 10,000+ units — laying policy groundwork for industry-wide scaling. This is institutional, government-level action rather than a single company's claim — against a backdrop of China's embodied AI funding already exceeding 90 billion yuan in H1, official policy is elevating "deployment" from corporate narrative to a national-level agenda. While it includes forward-looking targets, given its industry-wide influence this counts as today's major policy event.Source: Huaxing Low-Altitude and Robotics source (WeChat, CN)
Reports: Tesla's Third-Gen Optimus Nears Finalized Design, Musk Sets Hard Production Targets · humanoid
According to sources including the "Humanoid Robot Alliance," Tesla's third-generation Optimus is nearing a finalized design, with Musk reportedly warning he'll replace the entire procurement team if production targets aren't met; the Fremont factory that once built the Model S/X has reportedly been converted to a humanoid production line. Separately, Citizens initiated coverage of Tesla at "Market Perform," warning that revenue from Optimus and Robotaxi will fall short of Wall Street expectations, with other analysts also skeptical of Tesla's 2027–2028 scale-up timeline.⚠️ Single-source claim (the finalized design and personnel remarks are rumors/reported statements)Source: Humanoid Robot Alliance source (WeChat, CN)
Mitsubishi Motors Partners With a Startup on Humanoid Robots, Targets 2027 Production · humanoid
According to the Yomiuri Shimbun and others, Mitsubishi Motors is partnering with a startup to develop factory humanoid robots, targeting 2027 production and initial deployment in its own factories — extending the "automaker builds humanoids" pattern (paralleling Tesla and Hyundai's Boston Dynamics).⚠️ Planning-stage claimSource: Netauto source
US Regulators Warn Robotaxis Are "Functionally Deficient" as Federal and State Rules Diverge · autonomy
US regulators have warned of public safety risks from robotaxis being "functionally deficient," even as the industry rapidly expands (Waymo's fleet is nearing 4,000 vehicles). Within the same window, rules are diverging: NHTSA's administrator has signaled plans to remove the mandatory-steering-wheel requirement for autonomous vehicles, loosening design constraints, while New Jersey is moving to ban pure-vision approaches. Just as global robotaxi expansion abroad gained a unified framework via UN L4 technical regulations, regulatory compasses across individual jurisdictions remain fragmented.Source: STCN (Shanghai Securities News) source (WeChat, CN)
Robotics IPO Wave on the Hong Kong Stock Exchange: Yikong Zhijia and Rokae List on the Same Day · industrial
Autonomous mining truck company Yikong Zhijia listed on the Hong Kong Stock Exchange (over 2,500 units cumulatively deployed, annual revenue around 1.4 billion yuan), and industrial robot maker Rokae listed on the HKEX main board the same day; alongside Unitree's Sci-Tech Innovation Board IPO registration taking effect, the robotics/autonomous-driving listing window remains hot at both fundraising tiers.Source: Phoenix Finance source · Gasgoo source
Dexmal Turns Robot Foundation Models Into MaaS: 1 Yuan per Million Tokens · adjacent
Dexmal has launched an embodied MaaS (Model-as-a-Service) offering and DexOS, pricing its robot foundation models by token at rates as low as 1 yuan per million tokens, aiming to lower the barrier for models to scale into real-world deployment — a commercialization play that shifts the embodied "brain" business from selling hardware to selling model services.Source: Robot Frontier source (WeChat, CN)
LG Doubles Down on Physical AI, Developing Its Own Physics-Aware World Model · world-model
Korean media report that LG is making a major bet on Physical AI, developing its own world model that "understands physical laws," seen as directly competing with Nvidia in world models; LG Electronics has also begun early-stage mass production of joint actuators for humanoid robots. This continues the pattern of South Korean conglomerates building out both manufacturing and intelligence layers for physical AI.Source: Seoul Economic Daily source
Hardware · Supply Chain
· LG Actuators: LG Electronics begins early-stage mass production of joint actuators for humanoid robots, staking out a position in a core humanoid drivetrain component source
· Mobilint REGULUS NPU: targeting year-end mass production of its second-generation robot inference chip source
· Dexterous hand costs: an industry chain report estimates dexterous hands account for roughly 14%–18% of a full humanoid robot's cost, directly determining how many use cases it can enter source (WeChat, CN)
· Dmetacom Robotics: China Telecom has taken an equity stake in the firm, which focuses on tactile dexterous hands — telecom capital moving into the upstream tactile-sensing sector source
Top comments (0)