Learning Generalizable and Well-Shaped Reward Functions from Too Few Demonstrations

Ziyi Liu, Grace Zhang, Gaurav S. Sukhatme

Keywords: Inverse Reinforcement Learning, Imitation Learning, Learning from Few Demonstrations

TL;DR: We learn a reward function and policy from a few task demonstrations in an environment with variations using a multi-task demonstration dataset.

Abstract: Inverse reinforcement learning (IRL) is an important problem that aims to learn a reward function and policy directly from demonstrations, which can often be easier to provide than a well-shaped reward function. However, many real-world tasks include natural variations (i.e., a cleaning robot in a house with different furniture configurations), making it costly to provide demonstrations of every possible scenario. We tackle the problem of few-shot IRL with multi-task data where the goal is for an agent to learn from a few demonstrations, not sufficient to fully specify the task, by utilizing an offline multi-task demonstration dataset. Prior work utilizes meta-learning or imitation learning which additionally requires reward labels, a multi-task training environment, or cannot improve with online interactions. We propose Multitask Discriminator Proximity-guided IRL (MPIRL), an IRL method that learns a generalizable and well-shaped reward function by learning a multi-task generative adversarial discriminator with an auxiliary proximity-to-expert reward. We demonstrate the effectiveness of our method on multiple navigation and manipulation tasks.

Primary Area: reinforcement learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 13161