HY-VLA

HY-VLA (Hy-Embodied-0.5-VLA) is an end-to-end Vision-Language-Action system developed by Tencent Robotics X × Tencent Hy Team. It spans the full robot learning stack: data collection, model design, pre-training, SFT, RL post-training via FlowPRO, and real-world deployment. Built on the Hy-Embodied-0.5 MoT backbone, HY-VLA integrates a flow-matching action expert, a compact memory encoder for multi-frame history, and a delta-chunk action representation decoupled from embodiment-specific kinematics. Powered by 10,000+ hours of UMI demonstrations collected with optical motion-capture, achieving 90.9%/90.1% on RoboTwin 2.0 benchmark. Demonstrates robust cross-embodiment transfer across four real-world platforms.