Dexterous Manipulation through Imitation Learning: A Survey

Authors: Shan An, Ziyu Meng, Chao Tang et al. Submitted: April 2025 (v5: December 2025) Venue: arXiv 2504.03515

Abstract

Traditional computational methods struggle with the complexity of multi-finger control in unstructured settings, while trial-and-error approaches demand substantial data and careful tuning. Imitation learning offers an alternative: robots acquire fine-grained coordination and contact dynamics directly from human examples, avoiding extensive simulation or manual reward engineering.

The survey synthesizes current research on imitation learning for dexterous manipulation, identifies emerging challenges, and proposes future research directions.

Key Topics Covered

• Multi-finger control learning from demonstrations
• Data collection methods: teleoperation, kinesthetic teaching, motion capture
• Policy representation: behavior cloning, diffusion policies, energy-based models
• Contact dynamics and fine-grained manipulation
• Generalization to novel objects and environments