A Minimalist Retargeting-Guided Reinforcement Learning Recipe for Dexterous Manipulation

TL;DR

REGRIND is a minimalistic reinforcement learning framework that enables dexterous manipulation on robotic hands using a single human demonstration. It retargets human hand-object motion to robot kinematics while preserving spatial and contact relationships, then trains a residual policy in simulation to track object-centric keypoints. The policy transfers zero-shot to real hardware after system identification, achieving human-like tool use.

Why It Matters

This approach simplifies the development of dexterous manipulation skills without requiring extensive training data or complex reward engineering. It enables rapid deployment of contact-rich tasks on diverse robotic platforms, reducing the barrier to entry for real-world robotic manipulation. The insights into sim-to-real transfer provide actionable guidance for future research in contact-rich control.

Key Findings

• Retargeting human hand-object motion while preserving contact and spatial relationships enables effective policy learning in simulation.
• A residual RL policy trained on object-centric keypoints achieves zero-shot transfer to hardware with minimal domain adaptation.
• System identification and careful modeling of contact dynamics are critical for successful sim-to-real transfer in dexterous manipulation.