Friday, June 21, 2019, 11:00
Room 02-012, Georges-Köhler Allee 102, Freiburg 79110, Germany
We are working on the implementation of model predictive control on legged robot, i.e. high-dimension underactuated articulated dynamical systems. While there is a lot of interest of the community about model-predictive control in this context, such an approach is yet not able to beat the state of the art, based on heuristic-based controllers. The purpose of the talk is to explain what are the particular difficulties that are preventing us (at least for now) to implement off-the-shelf methods to control this class of systems. In a first part, we will explain how the system is modeled and what are the various optimal-control problems can be written to describe locomotion tasks. In a second part, we will discuss the transcription we are using to convert the selected optimal-control problem into a static optimization algorithm.
In particular, we will insist on how we are handling constraints. The last part will focus on globalization issues, i.e. what are the nonconvexity properties of the considered static optimization problem, and what are their impacts. Unfortunately, nonconvexity is a key property when considering locomotion problems. For example, making a locomotion step (breaking a contact and creating a new contact) is often in another convex region of the optimization problem, hence is unlikely to be discovered by the solver, or only with large computation cost. Of course, we have to pay the cost at each new step. We propose an approach to mix reinforcement learning approaches with model predictive control, and rely on caching the past motion solutions in a "memory of motion" to enable globalization properties of the solver.