Progress and prospects in optimization for control and learning in robotics

Justin Carpentier (Inria / ENS Paris)

Over the past decades, optimisation has become a core component of robotics. It provides a consistent and flexible way to formalise and solve complex problems in robotics, ranging from simulation to learning and control. While writing optimisation solvers was initially the job of specialists in optimisation, the past few years have seen the emergence of new solvers developed by roboticists, and for robotics. This paradigm change corresponds to a turning point in the community, justified by the absence of off-the-shelf solvers capable of precisely and efficiently handling the specific features of robotic problems. In this talk, I will present a general overview of our recent contributions to optimisation for robotics. In the first part, I will first review contributions around the development of generic and efficient solvers to tackle a wide variety of optimisation problems arising in robotics, ranging from standard quadratic problems (QPs) to nonlinear problems (NLPs), trajectory optimisation (TO) and model predictive control (MPC) problems. The second part will deal with physical simulation. I will highlight recent progress to improve simulator accuracy and efficiency and draw some perspectives about differentiating physics, which is by nature non-smooth.