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

EmbodiSkill introduces a training-free framework enabling embodied AI agents to autonomously improve their skills through reflection on task execution trajectories. By distinguishing between skill deficiencies and execution lapses, the system allows frozen language models to achieve significantly higher task success rates, with a Qwen 3.5-27B model reaching 93.28% success on ALFWorld benchmarks.

EmbodiSkill represents a meaningful advance in embodied AI by addressing a fundamental challenge: how agents learn from failures in dynamic, physical environments. Unlike digital settings where skill refinement is straightforward, embodied agents must navigate variable layouts and object states, making it difficult to identify whether poor performance stems from incorrect skills or temporary execution errors. The framework's innovation lies in its skill-aware reflection mechanism, which analyzes trajectories relative to current skills and selectively updates only when genuine skill gaps exist.

The research builds on growing recognition that large language models can serve as powerful executors when augmented with procedural knowledge. Rather than requiring expensive retraining, EmbodiSkill operates as a training-free wrapper that accumulates domain-specific guidance from the agent's own interactions. This approach mirrors how humans refine learned procedures through deliberate practice and error analysis.

The reported results merit attention within the AI research community. A frozen Qwen 3.5-27B model achieving 93.28% task success substantially outperforms unaugmented GPT-4.2, suggesting that skill accumulation multiplies the effectiveness of base models. This has implications for building deployable embodied systems without constant model retraining, reducing computational costs while improving real-world performance.

Looking forward, validation across additional embodied benchmarks and real-world robotic platforms will determine whether this approach generalizes beyond academic datasets. The framework's applicability to multi-agent scenarios and its scalability to longer task horizons remain open questions. Success in these areas could establish skill-aware reflection as a standard technique in embodied AI development.

• →EmbodiSkill enables embodied agents to self-evolve skills without retraining by distinguishing skill gaps from execution lapses
• →A frozen Qwen 3.5-27B model achieved 93.28% task success on ALFWorld, outperforming unaugmented GPT-4.2 by 31.58%
• →The training-free framework reduces computational overhead while improving embodied agent performance through trajectory-based reflection
• →Skill-aware reflection preserves valid guidance while updating only genuinely deficient skill components
• →The approach demonstrates that procedural knowledge accumulation enhances large language model effectiveness in dynamic environments