Beyond adversarial examples: sampling and repairing diverse failures with RADIUM
Keywords: failure prediction, test-case generation, adversarial optimization
TL;DR: Our method predicts a diverse set of safety-critical failure scenarios for learning-based autonomous systems, then repairs the control policy to reduce the severity of those failures.
Abstract: Recent years have seen large numbers of learning-enabled autonomous systems deployed in the real world. Unfortunately, increased deployment has seen a corresponding increase in accidents involving these systems. We must be able to predict the ways in which these systems might fail and take steps to prevent those failures \textit{before} deployment. Existing tools for failure prediction struggle to search over high-dimensional environmental parameters and provide little guidance on how to mitigate failures once they are discovered. In this paper, we develop a novel framework to efficiently predict failures and propose policy parameter updates to mitigate those failures. By re-framing adversarial optimization as a sequential inference problem, our approach is able to generate a more diverse set of challenging failures, which in turn lead to more robust repaired policies. We propose both gradient-free and gradient-based approaches to solving this inference problem, achieving state-of-the-art performance for policy repair, and we include a theoretical and empirical evaluation of the trade-offs between the two.
Supplementary Material: zip
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.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
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: 3940
Loading