Event Location: NSH 1507

Abstract: obust to both external disturbances and modeling error. We describe a walking controller that functions by coordinating multiple low-dimensional optimal controllers. We break a simplified model of the dynamics into several subsystems that have limited interaction. Each of the subsystems are augmented with coordination variables and we use a Dynamic Programming algorithm to generate optimal controllers for the augmented subsystems. We then use value functions to coordinate the augmented subsystems by managing tradeoffs of the coordination variables, producing an optimal controller for the simplified dynamic model. Inverse dynamics are then used to generate joint torques for the full rigid-body model of the robot. In a simulation based on the Sarcos robot, we demonstrate the robustness of this method to unexpected external disturbances such as pushes (both impulsive and continuous), trips, ground elevation changes, slopes, and regions where it is prohibited from stepping.

When implementing this controller on physical hardware (the Sarcos humanoid robot), we must also cope with significant modeling error. We produce stable walking by modifying our controller to include individual joint PD gains as well as modifying the swing leg subsystems to include acceleration as a state (and control jerk rather than acceleration). We also present two modifications to the Dynamic Programming algorithm, a multiple-model variant and a learning- based variant, that allow us to generate policies that are more tolerant of modeling error.

Committee:Christopher G. Atkeson (Chair)

J. Andrew (Drew) Bagnell

Hartmut Geyer
Jerry Pratt, Institute for Human and Machine Cognition