EmbodiedEvo: Self-improving Coding Agents for Physical Intelligence

TL;DR

EmbodiedEvo introduces a coding agent framework that learns reusable skills and lessons through iterative self-improvement in a simulated environment, enabling efficient adaptation to new physical tasks. It outperforms existing methods on RoboDojo-Sim, especially in memory-intensive and procedurally coordinated tasks, while reducing execution costs significantly.

Why It Matters

This approach advances physical AI by enabling agents to autonomously refine and transfer knowledge across tasks, reducing reliance on extensive real-world training and improving efficiency in complex robotic workflows.

Key Findings

• EmbodiedEvo achieves a 44.48 average score and 37.82% success rate on RoboDojo-Sim, exceeding all verified prior results.
• Skills and Lessons enable effective reuse of coordinated procedures and improve performance on memory-dependent tasks, with stronger models enhancing weaker executors by up to 8.86 percentage points.
• Using learned Skills and Lessons reduces execution cost by 53.0% compared to raw interaction history while maintaining comparable task performance.