Explore, Execute, Evolve: A Skill Acquisition and Reuse Loop for Embodied Agents

2026-09-29

AI Summary

This study addresses the challenges of generalizing to unseen tasks and the high cost of exploration from scratch for embodied agents by proposing RoboSkill, a framework that establishes an "exploration-execution-evolution" closed loop for skill acquisition and reuse. Methodologically, it integrates tactile feedback to reduce interaction uncertainty and introduces reusable code-augmented textual guidance to lower inference overhead. Furthermore, it synergizes multimodal agents, vision-language-action models, and world-action models to drive the dynamic evolution of a skill library. Experiments demonstrate that RoboSkill improves success rates by 12.5%–25.0% and reduces execution time by 7.6%–72.4% on the LIBERO-10 benchmark. In real-world deployment, it achieves an 8.3% increase in success rate while reducing task completion time by at least 14.4%.

Abstract

Vision-language-action and world-action models have demonstrated impressive capabilities in robotics, yet generalization to unseen tasks remains challenging. More recently, general-purpose multimodal agents have shown great potential for zero-shot robotic task solving. However, they often incur high execution costs by reasoning and exploring the physical world from scratch. To reduce these costs, we introduce RoboSkill, a framework that connects skill acquisition and reuse through an Explore, Execute, Evolve loop. Within this loop, the agent explores to gather task-relevant information, executes tasks while adapting to feedback, and evolves its skill library based on execution records. It then reuses these skills to guide exploration and execution in the next cycle, closing the loop. To improve loop efficiency, we complement vision with tactile feedback to reduce uncertainty during physical interaction. We further augment textual guidance with reusable code to reduce reasoning overhead during skill reuse. On LIBERO-10, RoboSkill improves first-episode success rates by 12.5--25.0 percentage points and reduces average runtime by 7.6--72.4% across four agents. On real robots, it improves success rates by 8.3 percentage points and reduces average runtime for successful trials by at least 14.4%.

Problem

• embodied agents
• skill acquisition
• generalization
• execution cost
• robotics

Innovation

• Skill Acquisition and Reuse
• Embodied Agents
• Tactile Feedback
• Code Augmentation
• Explore-Execute-Evolve Loop