LCSR Seminar: Learning Complex Robotic Behaviors with Optimal Control

Description

Ludovic Dominique Righetti, a professor in electrical and computer engineering and in mechanical and aerospace engineering at New York University and an international chair at the Artificial and Natural Intelligence Toulouse Institute, will give a talk titled "Learning Complex Robotic Behaviors with Optimal Control" for the Laboratory for Computational Sensing + Robotics.

Abstract:

Model-predictive control (MPC) and reinforcement learning (RL) have successfully been used to generate complex robotics behaviors, from flying robots to humanoids. While in principle they can both solve the same optimal control problems, their real-world performance can differ dramatically across tasks. Furthermore, both techniques remain largely unsuccessful for complex manipulation problems. In this talk, I will revisit the sometimes-surprising optimization principles underlying these methods and discuss how this can help us design more reliable solvers, especially when working with sensor-based learned world-models. Throughout the talk, our recent results at the intersection of MPC, RL, and world-model learning for a variety of legged locomotion and manipulation tasks will illustrate these ideas. In particular, I will argue that as learned models become more complex, optimization algorithms need to increase in simplicity and will point to existing challenges at the intersection of data and computational efficiency to solve complex locomotion and manipulation tasks. Since the algorithms we design are intended for real applications that could change how we organize our societies, I will end the presentation with a broader discussion on the impacts of robotics research on society and the role engineers ought to play.