Task-agnostic Lifelong Robot Learning with Retrieval-based Weighted Local Adaptation

TMLR Paper5893 Authors

Abstract: A fundamental objective in intelligent robotics is to move towards lifelong learning robots that can learn to manipulate in unseen scenarios over time. However, continually learning new tasks and manipulation skills from demonstration would introduce catastrophic forgetting due to data distribution shifts. To mitigate the problem, we store a subset of demonstrations from previous tasks and utilize them in two manners: leveraging experience replay to retain learned skills and applying a novel Retrieval-based Local Adaptation technique to recover relevant knowledge. Besides, task boundaries and IDs are unavailable in scalable, real-world settings, our method enables a lifelong learning robot to perform effectively without relying on such information. We also incorporate a selective weighting mechanism to focus on the most ''forgotten'' action segment, ensuring effective skill recovery during adaptation. Experimental results across diverse manipulation tasks demonstrate that our framework provides a plug-and-play paradigm for lifelong learning, enhancing robot performance in open-ended, task-agnostic scenarios.

Submission Length: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Adam_M_White1

Submission Number: 5893