Generalization ability under extreme data efficiency
- Only about 2 hours of data collection, generalizing to thousands of objects and various environments
- Data usage efficiency improved by 250 times compared to Figure
- Reinforcement learning-driven, quickly adapts to changes in new objects, lighting, stacking, etc.
Dexterous grasping and complex task reasoning collaboration
- Supports language instructions, multi-target automatic sorting, and CoT-style long-horizon planning
- Achieves stacked object retrieval, pose closed-loop correction, and repeated grasping
- Can utilize the environment for non-grasping actions, breaking traditional grasping limits
Safety alignment and efficient simulation transfer
- SafeVLA model achieves an 83% improvement in human-robot interaction safety
- Based on CMDP and multi-disturbance simulation environments, enhances anti-interference capability
- Full-stack support for Sim-to-Real transfer, efficient deployment in real-world scenarios