Policy Stitching: Learning Transferable Robot Policies

TL;DR

Policy Stitching introduces a modular transfer learning framework for robots, enabling rapid adaptation to new tasks and configurations by stitching separately trained modules, demonstrated through superior zero-shot and few-shot transfer results.

Contribution

It proposes a novel modular policy framework with latent space alignment, allowing direct stitching of trained modules for efficient transfer learning in robotics.

Findings

Superior zero-shot transfer performance in simulated and real-world tasks

Effective few-shot adaptation with minimal additional training

Outperforms existing transfer learning methods in complex manipulation tasks

Abstract

Training robots with reinforcement learning (RL) typically involves heavy interactions with the environment, and the acquired skills are often sensitive to changes in task environments and robot kinematics. Transfer RL aims to leverage previous knowledge to accelerate learning of new tasks or new body configurations. However, existing methods struggle to generalize to novel robot-task combinations and scale to realistic tasks due to complex architecture design or strong regularization that limits the capacity of the learned policy. We propose Policy Stitching, a novel framework that facilitates robot transfer learning for novel combinations of robots and tasks. Our key idea is to apply modular policy design and align the latent representations between the modular interfaces. Our method allows direct stitching of the robot and task modules trained separately to form a new policy for fast adaptation. Our simulated and real-world experiments on various 3D manipulation tasks demonstrate the superior zero-shot and few-shot transfer learning performances of our method. Our project website is at: http://generalroboticslab.com/PolicyStitching/ .