Dissertation Talk: Optimizing for Robot Transparency
Seminar: Dissertation Talk: EE | May 3 | 3-4 p.m. | 250 Sutardja Dai Hall
As robots become more capable and commonplace, it is increasingly important that the policies they execute are transparent. For instance, engineers should have an idea of which situations their robot may act incorrectly in, and end-users should be able to anticipate how a robot they are interacting with will behave in various situations. This is essential for building trust, enabling seamless human-robot collaboration, verifying what a robot has learned, and deploying robots in safety-critical situations. Unfortunately, passive familiarization to robot policies takes a while -- it would take someone hours, possibly days, of riding in an autonomous car before they understand its driving style and which situations it can and cannot handle. I will talk about how to speed up this process of making policies transparent, by optimizing for informative examples to show end-users, that enable them to more quickly gain an understanding of how a robot will act, when it will fail, and why it failed.