Numerical Optimal Control in Agile Robotics

Prof. Jonas Buchli, Dr. Diego Pardo and Michael Neunert

ETH Zurich

Wednesday, December 07, 2016, 10:30 - 12:30

Room 01-012, Georges-Köhler-Allee 102, Freiburg 79110, Germany


10:30 - 10:50 Optimal and learning control for agile and versatile robots ​

Speaker: Jonas Buchli, (Michael Neunert, Diego Pardo), ETH Zürich

Abstract: We will present a unified approach to learning and optimal control applied to robotic walking, running, flying, and manipulation. A key aspect is addressing generation of behavior as an optimal control and planning challenge. The resulting problems can be solved by a pipeline of iterative optimal and learning control techniques, based on nonlinear programming, sequential quadratic control and path integral reinforcement learning. In this series of talks, we will first show the common mathematical background of these methods that allow for a unified view. Then we will emphasize two core techniques developed from this background: iterative nonlinear optimal control and direct transcription methods. 

11:00 - 11:45 Efficient Iterative Nonlinear Optimal Control

Speaker: Michael Neunert, ETH Zürich

Abstract: In contrast to traditional robotic setups that operate in prepared environments, mobile robots need to constantly adapt and react to their environment. Additionally, to use the full potentials of these robots, we need to design controllers that reason about the system's dynamics and possibly about contact forces when interacting with the environment. These two requirements lead to a high dimensional, nonlinear planning and control problem. We tackle this problem by using iterative nonlinear optimal control techniques, which sequentially approximate the nonlinear problem as linear quadratic control subproblems, which can be solved in closed form. This allows us to predict and design trajectories for long time horizons within a few milliseconds without simplifying the system dynamics. We have demonstrated how the approach can be applied as a nonlinear model-predictive controller to ground and flying robots. Furthermore, when applied to walking robots, the controller automatically discovers contact timings and sequences or even full gaits.

12:00 - 12:30 Direct Methods for Trajectory Optimization on Legged Robots

Speaker: Diego Pardo, ETH Zürich

Abstract: Legged robots are non-linear, hybrid, and underactuated systems under the permanent influence of contact forces. This complexity limits the capability of conventional approaches for producing skillful solutions to motion tasks as fundamental as walking.  In this talk, I present a new trajectory optimization method that searches for solutions within the constraint-consistent subspace defined by the robot’s contact configuration. This research deals with the complexity associated with optimizing over the whole-body dynamics of a floating-base system subject to contact constraints. The results of this project  show that feasible motions, transferable to the real system, are obtained when the hybrid nature of the system is considered and appropriately used in an optimal control formulation.