国内首个通过具身智能国标测试VLA模型开源

Core Insights - The Beijing Humanoid Robot Innovation Center has officially launched the XR-1, the first and only embodied VLA model in China to pass the national standard test for embodied intelligence, marking a significant advancement in the domestic embodied intelligence industry [1][3] Group 1: XR-1 Model - The XR-1 model addresses the industry pain point of the disconnect between "visual perception" and "action execution" by utilizing three core capabilities: cross-data source learning, cross-modal alignment, and cross-ontology control [3] - The model employs the innovative UVMC technology to create a mapping bridge between vision and action, enabling robots to have "instinctive responses" and autonomously handle complex environments and emergencies [3] - In practical applications, robots equipped with XR-1 can perform complex tasks such as passing through five doors consecutively and accurately sorting materials without human intervention [3] Group 2: Data Support and Upgrades - The RoboMIND 2.0 dataset has been comprehensively upgraded, increasing operational trajectory data to 300,000 entries, covering 11 industrial, commercial, and household scenarios, and adding 12,000 entries of tactile operation data to support long-range collaborative task training [5] - The ArtVIP dataset offers over 1,000 high-fidelity digital twin articulated objects, with simulation data combined with real machine data, resulting in an average success rate improvement of over 25% for robot tasks [5] - Both datasets provide robust data support for the XR-1 model, significantly lowering the training threshold for the model [5]