Optimizing Locomotion: Learning Control at Intermediate Reynolds Number

Russ Tedrake
MIT

Abstract

In this talk I'll describe our efforts in using computational tools from optimal control theory (including machine learning and motion planning algorithms) to design efficient and agile control systems for locomotion. In particular, I'll emphasize new results applying these ideas to bird-scale aerial vehicles in complicated fluid regimes. Fluid dynamics, particularly at intermediate Reynolds numbers, represents one of the hard sciences where experiments still have a clear advantage over theory. It also happens that this dynamic regime offers some incredibly exciting controls problems; problems where classical control approaches have not made significant progress. I will argue that optimal control methods based on approximate models and model-free reinforcement learning methods are very well-suited to these regimes and may be the most natural route to finding efficient, high-performance control solutions. I'll describe our learning experiments with robotic birds (which fly with flapping wings) and with an airplane that can land on a perch. I will also briefly describe how these tools can be applied naturally to the control of minimally-actuated walking machines on rough terrain.

Russ Tedrake is the X Consortium Assistant Professor of Electrical Engineering and Computer Science at MIT, and a member of the Computer Science and Artificial Intelligence Lab. He 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. In 2008, he received an NSF CAREER award, the MIT Jerome Saltzer award for undergraduate teaching, and was named a Microsoft Research New Faculty Fellow.