RoboCopilot: Human-in-the-Loop Interactive Imitation Learning
Authors: Philipp Wu, Yide Shentu, Qiayuan Liao, Ding Jin, Menglong Guo, Koushil Sreenath, Xingyu Lin, Pieter Abbeel Submitted: March 2025
Overview
Addresses the limitation of passive demonstration collection (record-then-train) by building an interactive system where a human teacher and an autonomous robot policy share control of a bimanual robot. The human can seamlessly intervene, correct, and hand back control during task execution.
Key Design
- Bilateral teleoperation interface: compliant, force-reflecting — human feels what the robot feels
- Dynamic control switching: human ↔ autonomous policy transitions happen in real-time without stopping the task
- Interactive imitation learning: every human intervention becomes a training signal for the policy
Why This Matters
Traditional LfD:
- Collect all demonstrations passively
- Train policy offline
- Deploy and evaluate
RoboCopilot approach:
- Deploy initial policy
- Human intervenes when policy fails
- Intervention becomes new training data
- Policy improves with each correction
This is DAgger (Dataset Aggregation) in hardware — the human teacher provides on-policy corrections, directly addressing the covariate shift problem.
Results
Validated on bimanual manipulation tasks in simulation and on hardware. Interactive teaching achieves faster skill acquisition than passive collection alone.