Hesai Technology has published its Spatial Intelligence White Paper, examining how spatial intelligence technology is evolving and where opportunities lie. The company says that as AI moves into the physical world, generating and delivering high-quality, multimodal 3D spatial data at scale becomes essential to advancing Physical AI. The paper also describes how Hesai's Kosmo combines integrated hardware and software to meet that need.

Leaderobot's Xingzhe R1 Pro is a full-stack embodied AI research platform designed to coordinate data, models, hardware, and application scenarios.
Zhongying Electronics (300327.SZ) has been exploring variable frequency drive solutions for servo and robot joint control over the past two years, and is evaluating deeper AI integration across its business segments. The projects remain in very small-batch trial and R&D design stages, with no change to the company's overall fundamentals.
Ant Group's robotics unit Robbyant has open-sourced LingBot-Touch, a tactile AI model for robot hands, releasing model weights and inference code. Trained on nearly 5,000 hours of data from three tactile sensor types, the model was evaluated on 26 real-world tasks involving fragile-object handling and two-hand coordination.
D-Robotics presented WM-LOCO, which jointly trains a recurrent world model and a PPO locomotion policy, using proprioception plus a single onboard depth image to give predictive context for foot placement. In simulation it reached 96.9%, 96.0%, 94.0% and 92.1% success on stepping stones, wide gaps, stairs and narrow balance beams. Across 40 real-world trials, a Unitree G1 humanoid achieved 92.5% success.

Ant Lingbo Technology has open-sourced LingBot-Touch, a tactile foundation model pretrained on nearly 5,000 hours of real-world data across magnetic, vision-based, and piezoresistive sensors, aiming to give robots reusable touch perception for fine manipulation.
Anthropic's new Fable 5.2 model reportedly achieved a 64.9% success rate in dual-arm robot tests, a result that helped spark a broad rally in Chinese robotics stocks on November 10.
Robbyant released LingBot-Touch, a tactile perception model built on what the company calls an embodied-native approach, together with an open-source technical report, model weights and code. The materials cover the project page, a technical paper on learning heterogeneous tactile representations for general dexterous manipulation, a GitHub repository and a Hugging Face collection. Robbyant says the model is designed around real-world tactile needs rather than lab-only settings.
LingBot-Touch ships in two open-source sizes, Tiny and Small, which trade representational capacity for lightweight deployment and target efficient, high-frequency tactile perception as contact conditions change quickly. Robbyant reports that as pretraining data grew to nearly 5,000 hours, the Small variant kept improving on tactile perception tests, which the company cites as evidence of scaling potential in tactile pretraining.
Robbyant introduced LingBot-Touch, an open-source tactile foundation model for dexterous manipulation, pretrained on nearly 5,000 hours of real tactile data. Through unified heterogeneous tactile modeling, it supports transfer across sensors, dexterous hands, tasks and action models. The company framed the release around a question: what changes when robots can feel, not just see. A demo clip accompanies the post.

Skild AI is inviting outside developers to contribute to Steel, the simulation framework it maintains as a shared foundation for diverse, scalable robotics simulation. The company said community contributions are key to unlocking scale, and that it is committed to keeping the project maintained. Code and a technical blog post were released alongside the call for contributions.
Skild AI has open-sourced a WebGPU implementation for testing and gaming, available in its Steel playground, where users move simulated objects with a mouse to complete robotic tasks. The release extends the company's open-source push around Steel, its in-house physics solver built with NVIDIA.

In its S1 blog post, Skild AI described simulation as scalable but limited in diversity. It says Steel is meant to fulfil the full potential of diverse, scalable simulation for robot learning.

Skild AI has open-sourced Steel, its in-house physics solver built with NVIDIA. The company says it runs simulations up to 13× faster than PhysX while using roughly 3× less GPU memory, aimed at making scalable simulation for manipulation broadly accessible.

Boston Dynamics mounted wheels on its humanoid Atlas to test how faithfully a reinforcement-learning whole-body controller performs with real actuators running at the edge. The company says the experiment is part of its push to make robots reliable and performant outside the lab. Wheels let the team stress-test controller behavior under real hardware constraints before trusting it in field conditions.

Apple's research team has unveiled HALO, a robot-brain preprint submitted to arXiv, sparking a rally in Chinese robotics stocks including Unitree and Qwen-linked names.
Apple's research team has disclosed progress on a robot "brain," helping strengthen the physical AI theme. Chinese robot-themed ETFs reportedly rose more than 2% collectively.
Apple researchers disclosed progress on a robot "brain," reporting an average 91% success rate in single-task tests. The news lifted physical-AI sentiment, with the Robot ETF South (159258) tracking index up over 3% intraday and more than 12% over 16 trading days.