On December 18, 2025, the Beijing Humanoid Robot Innovation Center took a significant step by officially open-sourcing XR-1. This model holds a unique position as it is the first and only embodied VLA (Vision-Language-Action) large model in China that has successfully passed the national standard test for embodied intelligence. Alongside XR-1, the center also open-sourced the supporting data foundation, RoboMIND 2.0, and the latest iteration of ArtVIP.

XR-1 stands out with its remarkable multi-scenario, multi-agent, and multi-task capabilities, coupled with a high degree of generalization. In terms of its technical underpinnings, it boasts three core competencies: the ability to learn across different data sources, align cross-modally (for instance, between visual and linguistic information), and control multiple agents. Moreover, it introduces a pioneering three-stage training paradigm that seamlessly integrates virtual and real-world data.