Today's Highlights
· ModelBest (Chinese AI lab spun out of Tsinghua) open-sources embodied model MiniCPM-Robot, crushing π0.5's memory-benchmark score of 10 with 53 points
· The founder of Huawei's embodied-brain division departs to start his own company; Cognibrain, positioned against JEPA, raises a nine-figure RMB round
· China's "Measures on Anthropomorphic AI Interaction Services" take effect, pulling emotional-companion robots into tight regulation
· Former head of DJI's robot vacuum business pivots to exoskeletons; Sequoia China and Shunwei Capital back Yueshan Innovation (OROX)'s angel round
· WAIC organizers disclose: China's core AI industry has surpassed RMB 1.2 trillion in scale
1. Research Progress
Universal transferable visual attack: a single adversarial patch defeats VLA models across the board · vla
As VLA (vision-language-action) models become the dominant "brain" for robots, the robustness of their visual input pipeline is emerging as a security weak point. This work by a Nanyang Technological University team, named a CVPR 2026 best paper, produced an adversarial patch that is universal and transferable across models — placed within a camera's field of view, it induces different VLA models' action policies to fail, with no need to tune it for each model individually. It's a reminder that when robots actually do physical work, "seeing one thing wrong" can lead to consequences far more severe than a misclassified image.
Nanyang Technological University team · CVPR 2026 best paper · Analysis: Ru-hulu的具身智能分享会 (Ru-hulu's Embodied Intelligence Sharing) source (WeChat, CN)
Open Source · Tools · Benchmarks
· MiniCPM-Robot (ModelBest × OpenBMB): ModelBest, a Tsinghua-linked AI lab, made its WAIC debut in embodied intelligence, open-sourcing two on-device models. The general-purpose VLA MiniCPM-RobotManip, trained on the multimodal on-device model MiniCPM-V 4.6, has just 1.5B parameters yet matches 3–4B-parameter models — by aggressively compressing visual tokens it packs historical observations into inference, achieving 1-minute native memory. On RMBench, which tests contextual memory, it scores 53 versus π0.5's 10 (near-random). At 60 frames of history, per-step inference drops from a traditional 125 TFLOPs to 3.3 TFLOPs, with single-step latency of 120ms on an H100 — about half that of π0.5. A second, 0.9B model, MiniCPM-RobotTrack (co-developed with Nanjing University), is built for offline local tracking, natively supporting Unitree's Go2 — it can track a person at 5+Hz even with the network cable unplugged, and posts scores of 89.8/73.4/80.4 across three tracking tasks, an open-source SOTA. Both are available on GitHub/Hugging Face. source (WeChat, CN)
· PhyAI (ModelBest): An embodied reasoning framework open-sourced alongside the models, built for fast on-device inference and large-scale cloud RL rollouts; the company claims speedups of roughly 2.85×/1.82×/2.28× for π0/π0.5/GR00T respectively on an RTX 5090. source (WeChat, CN)
2. Funding & Deals
Cognibrain (具脑磐石) | New round | Nine-figure RMB · world-model
The round was led by industrial capital with backgrounds in brain-inspired computing and embodied intelligence, with follow-on investment from existing shareholders; Duowei Capital served as exclusive financial advisor, and another financing round is closing in parallel. Founded in 2025, the company's founder and CEO Zhu Senhua was formerly "the top person for Huawei's embodied brain division," having served as director of Huawei Cloud's AI Algorithm Innovation Lab and led the Pangu embodied foundation model; he holds a PhD in cognitive neuroscience from the University of Pennsylvania. The team is pursuing a brain-inspired intelligence approach to build a "cognitive world model," aligned in direction with Yann LeCun's JEPA but emphasizing a full closed loop from "representation-prediction" through to "decision-execution," with a focus on low data requirements, high generalization, lifelong learning, and low power consumption. This marks another team with big-tech embodied-AI roots striking out independently, as the industry's buzzword shifts from "VLA" to "world models."Source: 36氪硬氪 (36Kr Hard Tech) source
Yueshan Innovation (OROX) | Angel round | Amount undisclosed · adjacent
Co-invested by Sequoia China and Shunwei Capital. Founder Chen Liangshen previously led DJI's supply chain and robot vacuum business, where he headed the launch of DJI's first sweep-and-mop robot, ROMO, which sold over 10,000 units in its first month. The company builds consumer-grade exoskeletons, applying DJI's "R&D–supply chain–distribution" playbook to a category still stuck between lab prototypes and mass production. A notable reference point: online retail sales of intelligent walking-assist exoskeletons in China rose 785% year-over-year from January to April this year, with retail prices falling from hundreds of thousands of yuan to the thousand-yuan range. This also reflects a shift in early-stage investing — teams with algorithm patents but no manufacturing and delivery capability are no longer the top pick.Source: 具身涌现 (Embodied Emergence) source (WeChat, CN)
Coram AI | Series B | $35 million · adjacent
Led by Ansa Capital and Battery Ventures. Coram builds AI video security software, connecting existing cameras to models for real-time understanding and alerts — a perception/monitoring-layer player adjacent to robotics.Source: ascendants.in source
Wildfire Robotics (山火机器人) | Angel round | Tens of millions of RMB · hardware
Led by Haiyuan Capital, with proceeds going toward team expansion and R&D. The company focuses on localizing production of core transmission components for humanoid robots in China, targeting the reducer/transmission segment — currently the tightest bottleneck upstream.Source: 链榜 (Chain Ranking) source (WeChat, CN)
3. Commercialization & Deployment
Honda launches commercial autonomous electric lawn mower, priced at $32,999 · autonomy
Honda has launched an autonomous electric lawn mower for the commercial market, priced at $32,999. Compared with passenger-carrying autonomous driving, lawn maintenance is a closed environment with low failure costs — putting L4 capability first into "no passengers, fault-tolerant" work machinery is a low-profile but genuinely scaling branch of autonomous-driving commercialization.Source: 新浪网 (Sina) source
Aomi Technology (OYMotion)'s dexterous-hand orders top 10,000 units in H1 · embodied
Dexterous-hand maker Aomi Technology (OYMotion) disclosed that orders surpassed 10,000 units in the first half of the year. In a WAIC week saturated with dexterous-hand "demos," this is one of the few figures that reflects real order volume; set against this year's price collapse for dexterous hands — from "100,000+ RMB" down to the "10,000 RMB" range — a rising volume number is worth more attention than spec sheets.Source: 新浪财经 (Sina Finance) source
San Francisco blackout briefly halts Waymo robotaxi service · autonomy
A major power outage in San Francisco briefly disrupted Waymo's robotaxi service and paused freeway routes, before service resumed following adjustments. Scaled autonomous vehicle fleets depend heavily on charging, dispatch, and communications infrastructure — a city-wide blackout serves as a real-world stress test of their redundancy and recovery capability.Source: 디지털투데이 (Digital Today) source
4. Industry Developments
Regulators convene closed-door meeting; emotional-companion robots pulled into tight regulation · humanoid
According to 21st Century Business Herald, regulators convened a closed-door meeting with leading companies last week on emotional-companion robots, focused on data security, cybersecurity, and standards for emotional interaction. The "Interim Measures on Anthropomorphic AI Interaction Services," effective July 15, has now formally taken effect, bringing all AI products/robots that provide sustained emotional interaction through anthropomorphic forms under tight regulation, requiring hard compliance measures including protections for minors, addiction-prevention features, content moderation, and extreme-emotion intervention. The category is red-hot: UBTech's U1 hyper-realistic humanoid sold more than 10,000 units in a single day, starting at RMB 119,800 and topping out at RMB 990,000, for an estimated transaction value of nearly RMB 2.2 billion (per brokerage estimates). But sustained companionship requires long-term collection of voice, imagery, and even private emotional records, magnifying data-leak and ethical risks — this "trillion-yuan opportunity" remains a chaotic category with unresolved gaps in data security, compliance, and user experience.Source: 21财经 (21st Century Business Herald) source
Official WAIC data: China's core AI industry surpasses RMB 1.2 trillion · adjacent
The 2026 World Artificial Intelligence Conference disclosed that China's core AI industry has surpassed RMB 1.2 trillion in scale, with "embodied intelligence going to work in factories" named as this year's most distinctive industry signal; organizers also stated that roughly 120,000 high-quality datasets have been built. Beyond the numbers, the notable shift is in narrative — a consensus emerged at this year's WAIC that the focus is moving from showcase demos to real deployment in factories and homes.Source: 新浪财经 (Sina Finance) source
Da Xiao Robotics unveils "Kaiwu World Model" Kairos 3.1 · world-model ⚠️ Vendor claim
Da Xiao Robotics (ACE Robotics) launched Kairos 3.1 (Kaiwu World Model 3.1) at WAIC, touting "native action integration" and claiming to be the world's first embodied world model driving a robot body via on-device execution. Both the "world's first" claim and the on-device real-time performance are the vendor's own statements; real-world generalization and deployment performance await third-party verification.Source: 新浪财经 (Sina Finance) source
Shouqi Ride Hailing launches SAIC robotaxi mass-production customization project, debuting in 2027 · autonomy
Shouqi Ride Hailing, together with SAIC Motor and a leading autonomous-driving technology provider, has launched a robotaxi mass-production customization project. Officials state it will be Shanghai's first factory-built, mass-produced, purpose-designed vehicle for high-level autonomous driving, with the new vehicle slated to debut in 2027. This remains at the planning stage and deployment is still far off, but an automaker directly defining a factory-built, mass-produced robotaxi is another example of OEMs bringing robotaxis into their core business.Source: 上观新闻 (Shanghai Observer) source
Google DeepMind: video generation models already "contain" the world models computer vision has been missing · world-model
Amid ongoing industry debate over the boundaries of "world models," Google DeepMind published an argument that today's video generation models already implicitly contain world models — the piece long missing from computer vision, they argue, can be "read out" of video generators. This is a research stance rather than a settled conclusion, but it resonates with this week's broader push from academia and industry to redefine "what counts as a world model," pulling the debate back from terminology toward the question of what these models have actually learned.Source: the-decoder.com source
Hardware · Supply Chain
· Weihong Co. / Changzhou Hanmu: Holds the top market share in dexterous-hand motors, with monthly shipments of around 18,000 units; reportedly a co-development partner for many leading dexterous-hand companies, with supply set to ramp up in the second half of the year. source (WeChat, CN)
· 9.9mm frameless torque motor: A company has built the world's smallest 9.9mm frameless torque motor, along with a 3.5g fingertip motor for dexterous hands — pointing toward hands with higher degrees of freedom and dimensions closer to a human hand's "fingertip heart." source (WeChat, CN)
· Dexterous-hand price drops: Robotic dexterous hands once priced at "100,000+ RMB" have now fallen into the 10,000 RMB range, and this price collapse is drawing in new entrants such as cross-border e-commerce players. source (WeChat, CN)
Top comments (0)