China's embodied intelligence sector achieved a major breakthrough on December 18: Beijing Humanoid Robot Innovation Center officially announced the open-source release of the embodied VLA large model XR-1. This is the first and only visual-language-action (VLA) large model in China that has passed the national standards for embodied intelligence, marking a key step for humanoid robots in China to move from "laboratory walking" to "performing tasks in real-world scenarios."

Robot Computer Office Artificial Intelligence

This open-source release includes not only the XR-1 large model but also a powerful data foundation RoboMIND2.0 and the latest version of the high-fidelity digital asset dataset ArtVIP. As the "cerebellum" of embodied intelligence, XR-1 has cross-body platform operation capabilities, enabling seamless migration of general operational knowledge across multiple robot platforms such as TianGong 2.0, UR, and Franka. With over a million self-collected body data, XR-1 performs well in seven generalization dimensions including object color, position, and background interference, and can accurately execute complex dual-arm skills such as picking, pushing, pulling, and rotating.

The Beijing Humanoid Robot Innovation Center has now built a complete "brain + cerebellum + body" ecosystem:

  • Physical Body: Relying on the "Embodied TianGong" platform, multiple types of bodies have been released, including TianGong 2.0 and TianYi 2.0;

  • Embodied Brain: Based on the "HuiSi KaiWu" platform, the WoW (I Understand) World Model and **Pelican-VL (Tianhu)** large model were previously open-sourced, responsible for high-level logical reasoning and task decomposition;

  • Embodied Cerebellum: The open-sourced XR-1 in this release is responsible for converting the brain's instructions into precise physical actions, achieving efficient collaboration between software and hardware.

AIbase analysis suggests that by fully open-sourcing core models and high-value datasets, the Beijing Humanoid Robot Innovation Center aims to lower industry development barriers, solve common challenges such as difficult data reuse and poor generalizability in embodied intelligence, and promote the domestic robotics industry into a new stage of large-scale application characterized by "full autonomy and better usability."