UBTECH on Tuesday unveiled the U1 Series, its first full-sized consumer humanoid robot, at a Shenzhen event under a new consumer brand called UWorld, TechNode reported. Founder Zhou Jian said the U1 has drawn more than 11,000 orders across sales channels since pre-orders opened on JD.com on June 2, with variants ranging from about $17,600 to $145,000 and shipments beginning Sept. 16. The robot is built around an emotional AI model designed for long-term interaction, and the company said user interaction data is encrypted and stored locally by default with no mandatory cloud uploads. The rollout is one of the largest consumer humanoid launches to date in China, putting emotional AI features and on-device personal data handling squarely under the Cyberspace Administration of China's generative AI rules as the format moves from apps into homes.
Read at TechNode ↗
NVIDIA opened new robotics roles in China's three main tech hubs across four teams: embodied AI, simulation, deployment and solutions architecture, TechNode reported. The embodied AI group is being staffed to work on dexterous manipulation, human body modeling from wearable sensors and whole-body control for next generation robotic systems, while the simulation team is tasked with building the training infrastructure that lets robots learn in virtual environments and transfer to the physical world. The deployment team will focus on humanoid robots and pushing algorithms into real-world scenarios. The recruiting drive extends U.S. physical AI R&D deep inside the mainland even as Washington's chip export controls continue to limit which of NVIDIA's GPUs can be sold there, sharpening the question for Commerce and BIS of how to draw the line between restricted hardware and unrestricted robotics know-how.
Read at TechNode ↗
Shenzhen-based DexForce closed multiple financing rounds worth more than 1 billion yuan ($140 million), pushing its post-investment valuation past 10 billion yuan ($1.4 billion) and setting the stage for an IPO, Pandaily reported. Investors include Shenzhen Capital Group, Qianhai Mother Fund, ICBC Capital and Lens Technology, whose founder Zhou Qunfei also personally contributed to the round. Founded in 2021 by Professor Jia Kui, the company sells the W1 Pro humanoid robot into industrial customers spanning automotive components, appliances and electronics, including Huawei and Midea, and projects 2026 revenue of 250 million to 300 million yuan ($35 million to $42 million) on close to 1,000 shipments. The milestone deepens Beijing's state and industrial capital bet on embodied AI, which the government has promoted as a domestic manufacturing upgrade priority.
Read at Pandaily ↗
A previously unknown Chinese AI team operating under the codename "MopMonk" hit a 73.1% success rate on CyberGym, the UC Berkeley cybersecurity benchmark selected for ICLR 2026, ranking it seventh globally and within striking distance of OpenAI's top tier models, Pandaily reported. CyberGym differs from standard evaluations in that it requires AI agentsAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise. to reason across full codebases and produce working proof-of-concept exploits for 1,507 real vulnerabilities drawn from open source projects in Google's OSS-Fuzz database. The team used Shanghai lab MiniMax's open weight M3 model as its foundation and layered on proprietary agent scaffolding for vulnerability memory management across long code paths, and it has published only a single technical report with no public track record. The result shows that open weight Chinese frontier models paired with agent scaffolding can now approach leading U.S. systems on offensive cyber tasks, a dual-use signal for U.S. export control designers still calibrating restrictions on inferenceInferenceRunning a trained AI model to answer a query, as distinct from training it. Inference is where most ongoing compute and energy demand now sits, and it is what chip export rules increasingly target alongside training. computeComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. and open weight modelOpen-weight modelAn AI model whose trained parameters are published so anyone can download, run and modify it, as opposed to one reachable only through the developer's API. The main policy tension is that open weights spread capability widely and cannot be recalled once released. releases.
Read at Pandaily ↗