PRISM: Personalized Refinement of Imitation Skills for Manipulation via Human Instructions

A team of researchers including Arnau Boix-Granell, Alberto San-Miguel-Tello, Magí Dalmau-Moreno, and Néstor García has developed PRISM (Personalized Refinement of Imitation Skills for Manipulation via Human Instructions), a novel approach that bridges Imitation Learning (IL) and Reinforcement Learning (RL) frameworks. The system starts with a generic imitation policy trained from user-guided demonstrations, then refines it through reinforcement learning where reward functions are generated iteratively from natural language task descriptions using the Eureka paradigm. This creates a seamless pipeline that transforms broad robotic skills into fine-grained behaviors.

What sets PRISM apart is its incorporation of human feedback during the refinement process. As the system generates intermediate rollouts, human operators can provide corrections that guide the policy adaptation to new goal configurations and constraints. This human-in-the-loop approach enables significant policy reusability and data efficiency. In simulated pick-and-place scenarios, PRISM demonstrated superior performance compared to policies trained without human feedback, showing improved robustness during deployment while simultaneously reducing computational requirements. The method represents a practical step toward more adaptable and efficient robotic systems that can learn complex manipulation tasks with less data and computational overhead.