EmbodiSkill: Skill-Aware Reflection for Self-Evolving Embodied Agents

Abstract: Embodied skills must evolve from execution trajectories, but failed tasks may reflect either incorrect skill content or an execution lapse where valid guidance was not followed. EmbodiSkill is a training-free framework that interprets each trajectory against the current skill, revises the skill body only when evidence warrants a change, and preserves valid guidance when failures arise from execution lapses. On ALFWorld and EmbodiedBench it consistently improves task success; a frozen Qwen3.5-27B executor reaches 93.28% on ALFWorld, outperforming GPT-5.2 used directly without skills by 31.58%.

@article{embodiskill,
  title={EmbodiSkill: Skill-Aware Reflection for Self-Evolving Embodied Agents},
  author={Ruofei Ju and Xinrui Wang and Xin Ding and Yifan Yang and Hao Wu and Shiqi Jiang and Qianxi Zhang and Hao Wen and Xiangyu Li and Weijun Wang and Kun Li and Yunxin Liu and Haipeng Dai and Wei Wang and Ting Cao},
  journal={arXiv preprint arXiv:2605.10332},
  eprint={2605.10332},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2605.10332},

  year={2026}
}