Early Career Spotlight Talk

Russ Tedrake
Massachusetts Institute of Technology
http://groups.csail.mit.edu/locomotion/russt.html

Biography: Russ is the X Consortium Associate Professor of Electrical Engineering and Computer Science at MIT, and a member of the Computer Science and Artificial Intelligence Lab. He is a recipient of the NSF CAREER Award, the MIT Jerome Saltzer Award for undergraduate teaching, the DARPA Young Faculty Award, and was named a Microsoft Research New Faculty Fellow.
Russ received his B.S.E. in Computer Engineering from the University of Michigan, Ann Arbor, in 1999, and his Ph.D. in Electrical Engineering and Computer Science from MIT in 2004, working with Sebastian Seung. After graduation, he joined the MIT Brain and Cognitive Sciences Department as a Postdoctoral Associate. During his education, he has also spent time at Microsoft, Microsoft Research, and the Santa Fe Institute.

Dynamic walking on rough terrain and flying like a bird: a
computational approach to exploiting nonlinear dynamics

In order for our robots to walk like a human or fly like a bird, we must overcome a few serious control problems. The dynamics of these systems are typically very nonlinear and underactuated, and are dominated by their rich, uncertain interaction with the environment. Furthermore, in order to locomote effectively, these robots are typically carrying relatively small actuators with limited power output. Most rigorous approaches to feedback design cannot address all of these complexities, and as a consequence, our best engineered systems to date move very conservatively, and often with hand-tuned or hand-designed controllers. In order to push our machines to the limits of their performance in terms of speed or agility (with actuators being pushed to saturation) or ultimate energy-economy, we need to be able to design control systems which can relinquish control to the nonlinear dynamics of the machine in order to accomplish a task.

In this talk, I'll describe a computational approach to this problem which builds on techniques from motion planning, nonlinear control and verification, system identification, and machine learning. Specifically, I will describe some modifications to randomized motion planning algorithms which allow them to plan effectively with serious dynamic constraints. I'll describe our extensions to tools from convex optimization for estimating regions of attraction which allow us to efficiently estimate invariant sets (or funnels) around stabilized trajectories, even allowing exact verification on systems with trigonmetric nonlinearities and impacts, and a motion planning algorithm which probabilistically fills the controllable space with provably safe feedback. Finally, I'll briefly describe our work on nonlinear system identification and on learning control inside the robust control framework. These ideas will be motivated and evaluated using examples of minimally-actuated walking robots moving over rough terrain and robotic birds that can land on a perch.