Skill-Space Shooting for Autonomous Robot Policy Improvement

TL;DR

The paper presents skill-space shooting, a method that enables autonomous robot policy improvement by using foundation models to explore corrective actions through reusable skills. These skills serve as a shared, learnable language for policy updates across tasks, reducing reliance on human demonstrations. Real-world experiments demonstrate sustained policy improvement without human intervention.

Why It Matters

This approach allows robots to autonomously learn from failures and adapt over time, making long-term deployment in dynamic environments more feasible. By reusing skills across tasks, it reduces the need for task-specific training data and human oversight, accelerating real-world adoption.

Key Findings

• Reusable skills act as a common language for corrective actions, enabling policy updates without human demonstrations.
• Foundation models can autonomously generate and evaluate skill sequences to improve task performance.
• Policy performance improves over multiple iterations in real-world settings, with skills transferable to new tasks.