Robohub.org
Todd Murphey: Active Learning in Robot Motion Control | CMU RI Seminar
Abstract: “Motion motivated by information needs can be found throughout natural systems, yet there is comparatively little work in robotics on analyzing and synthesizing motion for information. Instead, engineering analysis of robots and animal motion typically depends on defining objectives and rewards in terms of states and errors on states. This is how we formulate optimal control objectives, learning-based reward functions, and goals for sample-based planning. Sometimes coverage algorithms are used to manage the collection of data, distributing sensors across a domain, but often without explicitly reasoning about where useful information is likely to be present. This talk will focus on situations where motion is used for information, either obtaining information or communicating information. Ergodicity provides one means for relating a trajectory to information content, and I will talk about settings both in animal behavior and in physical Human-Robot Interaction where movement appears to be ergodic. These information-based analysis tools can also be used to synthesize motion, connecting control decisions to learning needs—a form of active learning. Examples from biology, vision-based tracking, and contact-based sensing will serve as examples throughout the talk.”
