Current browse context:
cs.RO
Change to browse by:
References & Citations
Computer Science > Robotics
Title: Task-Driven Detection of Distribution Shifts with Statistical Guarantees for Robot Learning
(Submitted on 25 Jun 2021 (v1), last revised 6 Nov 2023 (this version, v6))
Abstract: Our goal is to perform out-of-distribution (OOD) detection, i.e., to detect when a robot is operating in environments drawn from a different distribution than the ones used to train the robot. We leverage Probably Approximately Correct (PAC)-Bayes theory to train a policy with a guaranteed bound on performance on the training distribution. Our idea for OOD detection relies on the following intuition: violation of the performance bound on test environments provides evidence that the robot is operating OOD. We formalize this via statistical techniques based on p-values and concentration inequalities. The approach provides guaranteed confidence bounds on OOD detection including bounds on both the false positive and false negative rates of the detector and is task-driven and only sensitive to changes that impact the robot's performance. We demonstrate our approach in simulation and hardware for a grasping task using objects with unfamiliar shapes or poses and a drone performing vision-based obstacle avoidance in environments with wind disturbances and varied obstacle densities. Our examples demonstrate that we can perform task-driven OOD detection within just a handful of trials.
Submission history
From: Alec Farid [view email][v1] Fri, 25 Jun 2021 15:41:25 GMT (3683kb,D)
[v2] Mon, 20 Sep 2021 03:25:39 GMT (3705kb,D)
[v3] Tue, 28 Sep 2021 14:19:00 GMT (3705kb,D)
[v4] Tue, 26 Oct 2021 03:51:25 GMT (3705kb,D)
[v5] Wed, 1 Dec 2021 15:22:45 GMT (10376kb,D)
[v6] Mon, 6 Nov 2023 21:39:13 GMT (6287kb,D)
Link back to: arXiv, form interface, contact.