Summary
AutoSERL is a framework that leverages a single demonstration to fully automate the intervention process in real-world robot RL. Three complementary mechanisms work together: a sliding window intervention mechanism to guide exploration and prevent unsafe deviations, a safety recovery mechanism detecting and correcting failure states via predefined trajectory recovery points, and an intervention termination criterion that disables guidance once the policy can complete the task independently.
Key Contributions
- AutoSERL: single-demo automation of real-world robot RL intervention
- Sliding window intervention for safe exploration guidance without constant human presence
- Safety recovery mechanism with predefined trajectory recovery points
- Automatic intervention termination once policy achieves task competence
- Evaluated on 6 contact-intensive tasks (insertion, hanging, hinge-based) across 2 robot platforms
- Outperforms SERL with 20 demonstrations; matches HIL-SERL; 100% success on insertion tasks
- Improved robustness to positional variations
Significance
Reducing the demonstration requirement from 20 to 1 for real-world robot RL training is a major practical advance—AutoSERL makes SERL-style on-hardware RL accessible for rapid deployment in new tasks with minimal human effort.