Sample-Efficient Reinforcement Learning for Dexterous Policies

SERNF: Sample-Efficient Reinforcement Learning for Dexterous Manipulation

SERNF enables robots to improve dexterous manipulation skills through reinforcement learning with limited real-world interaction. By combining a normalizing-flow policy with a critic that evaluates sequences of actions, it supports stable fine-tuning of policies learned from demonstrations or simulation. We demonstrate the approach on tasks including scissors retrieval and tape cutting, in-hand cube rotation, and object pick-and-place.

For videos and further details, visit the external page SERNF project page.