EmbodiSkill: Training-Free Framework for Embodied Agent Skill Self-Evolution
EmbodiSkill is a framework that allows for the self-evolution of embodied skills without the need for training, utilizing skill-aware reflection and focused revision. It tackles the issue faced by embodied agents needing to evolve skills based on trajectories created during task performance. Current methods in digital settings often fall short, as unsuccessful attempts may indicate either flawed skill content or execution errors. EmbodiSkill evaluates each trajectory in relation to the existing skill, employs skill-altering evidence to modify the skill body, and facilitates improvement without further training. This framework caters to a variety of environments with different layouts, object conditions, and execution variables. The study is available on arXiv with the identifier 2605.10332.
Key facts
- EmbodiSkill is a training-free framework for embodied skill self-evolution.
- It uses skill-aware reflection and targeted revision.
- Existing skill self-evolution methods are mainly developed in digital environments.
- Failed task executions may reflect incorrect skill content or execution lapses.
- EmbodiSkill interprets trajectories with respect to the current skill.
- It uses skill-changing evidence to update the skill body.
- The framework is designed for diverse embodied environments.
- The research is published on arXiv (2605.10332).
Entities
Institutions
- arXiv