AED: Adaptable Error Detection for Few-shot Imitation Policy

Jia-Fong Yeh, Kuo-Han Hung, Pang-Chi Lo, Chi Ming Chung, Tsung-Han Wu, Hung-Ting Su, Yi-Ting Chen, Winston H. Hsu

Primary Area: applications to robotics, autonomy, planning

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.

Keywords: adaptable error detection, few-shot imitation, policy learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: The novel adaptable error detection (AED) problem is formulated for monitoring few-shot imitation policies' behaviors, and we propose PrObe to address the challenging problem by learning from the policy's feature representations.

Abstract: We study the behavior error detection of few-shot imitation (FSI) policies, which behave in novel (unseen) environments. FSI policies would provoke damage to surrounding people and objects when failing, restricting their contribution to real-world applications. We should have a robust system to notify operators when FSI policies are inconsistent with the intent of demonstrations. Thus, we formulate a novel problem: adaptable error detection (AED) for monitoring FSI policy behaviors. The problem involves the following three challenges: (1) detecting errors in novel environments, (2) no impulse signals when behavior errors occur, and (3) online detection lacking global temporal information. To tackle AED, we propose Pattern Observer (PrObe) to parse the discernable patterns in the policy feature representations of normal or error states. PrObe is then verified in our seven complex multi-stage FSI tasks. From the results, PrObe consistently surpasses strong baselines and demonstrates a robust capability to identify errors arising from a wide range of FSI policies. Finally, the visualizations of learned pattern representations support our claims and provide a better explainability of PrObe.

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: 4408