一种基于技能组合专家的具身智能持续学习方法及系统
By combining state-aware task decomposition and visual language model for combined skill semantic labeling, and integrating joint optimization training with execution-transformation dual expert modules, the problems of disjointed skill connections and unstable closed-loop execution in embodied intelligence continuous learning are solved, achieving the retention of old task capabilities and efficient learning of new tasks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN INSTITUTE OF TECHNOLOGY (SHENZHEN) (INSTITUTE OF SCIENCE AND TECHNOLOGY INNOVATION HARBIN INSTITUTE OF TECHNOLOGY SHENZHEN)
- Filing Date
- 2026-05-26
- Publication Date
- 2026-07-17
AI Technical Summary
Existing embodied intelligence continuous learning methods suffer from problems such as inconsistent skill connections, unstable closed-loop execution, and old skill parameters being easily disturbed by irrelevant updates in long-term operation tasks, making it difficult to effectively reuse fine-grained operation skills shared across tasks.
We adopt an embodied intelligence continuous learning method based on skill combination experts. We perform combined skill semantic labeling through state-aware task decomposition and visual language model to build a reusable skill library. We also introduce an execution-transformation dual expert module on the action decoding side of the pre-trained visual language action model to perform joint optimization training to enhance the hidden features of the action decoder.
It improves the continuity of skill transitions and the stability of closed-loop execution in long-term sequential tasks, reduces catastrophic forgetting, maintains the ability to perform old tasks, and enhances the learning efficiency of new tasks.
Smart Images
Figure CN122264029B_ABST