EmbodiSkill: Skill-Aware Reflection for Self-Evolving Embodied Agents
Researchers have introduced EmbodiSkill, a novel training-free framework designed to enhance the capabilities of embodied agents through skill-aware reflection and self-evolution. Unlike existing methods developed for digital environments that often produce coarse updates, EmbodiSkill addresses the complexities of physical settings by distinguishing between incorrect skill content and execution lapses. The framework interprets task trajectories to identify specific evidence for updating skill bodies or preserving valid guidance. Experimental results on ALFWorld and EmbodiedBench demonstrate significant performance improvements. Notably, when integrated with a frozen Qwen3.5-27B executor, EmbodiSkill achieved a 93.28% task success rate on ALFWorld, outperforming GPT-5.2 used as a direct agent by 31.58%. This study highlights the effectiveness of accumulating reusable procedural knowledge from agent trajectories, offering a robust solution for improving autonomy in diverse embodied environments without requiring additional model training.
Wire timeline
EmbodiSkill: Skill-Aware Reflection for Self-Evolving Embodied Agents
Researchers have introduced EmbodiSkill, a novel training-free framework designed to enhance the capabilities of embodied agents through skill-aware reflection and self-evolution. Unlike existing methods developed for digital environments that often produce coarse updates, EmbodiSkill addresses the complexities of physical settings by distinguishing between incorrect skill content and execution lapses. The framework interprets task trajectories to identify specific evidence for updating skill bodies or preserving valid guidance. Experimental results on ALFWorld and EmbodiedBench demonstrate significant performance improvements. Notably, when integrated with a frozen Qwen3.5-27B executor, EmbodiSkill achieved a 93.28% task success rate on ALFWorld, outperforming GPT-5.2 used as a direct agent by 31.58%. This study highlights the effectiveness of accumulating reusable procedural knowledge from agent trajectories, offering a robust solution for improving autonomy in diverse embodied environments without requiring additional model training.
cs.AI updates on arXiv.org