Interactive Imitation Learning for Dexterous Robotic Manipulation: Challenges and Perspectives
Authors: Edgar Welte, Rania Rayyes Submitted: May 2025 (v2: August 2025) Venue: arXiv 2506.00098
Abstract
Addresses dexterous manipulation in humanoid robotics, examining how robots can learn precise and adaptable control. Traditional learning approaches struggle due to:
- High-dimensional control spaces (21+ DOF hands)
- Limited training data
- Covariate shift (distribution mismatch between demonstration and deployment)
Surveys existing methods spanning imitation learning, reinforcement learning, and hybrid techniques. Emphasizes interactive imitation learning — where human feedback guides robot training — as a promising yet underexplored direction.
Key Concepts
- Covariate shift: demonstration data distribution ≠ deployment state distribution → compounding errors
- DAgger and variants: iterative correction by human teacher during robot execution
- Diffusion policies: model the full action distribution, robust to multimodality
- HITL: human-in-the-loop correction at test time improves long-horizon performance