Convex quadratic programming (QP) has become a major item in the robotics toolbox, with well-known applications including whole-body control, model predictive control (MPC), contact planning and state estimation. Current challenges when solving QP-formulated problems include feasibility (ensuring that a solution exists, e.g. when some problem parameters come from measurements), recursive feasibility (in MPC: ensuring the system does not steer towards unfeasible problems) and real-time performance. In this talk, we review new features brought by an upcoming generation of QP solvers. We will focus in particular on ProxQP, which can handle non-convex or unfeasible problems and always returns a principled solution. We will further develop how this enables the inclusion of differentiable QP layers in end-to-end trainable control pipelines.
|QP solvers benchmark
|ProxQP (includes QPLayer)
|PROXQP: an Efficient and Versatile Quadratic Programming Solver for Real-Time Robotics Applications and Beyond
|QPLayer: efficient differentiation of convex quadratic optimization
Stéphane is a research scientist at Inria. He received his M.Sc. in Computer science from the École Normale Supérieure (ENS Paris) in 2012, and his Ph.D. in Mechano-informatics from the University of Tokyo in 2016. After graduation, Stéphane has worked at CNRS as tenured researcher and at ANYbotics AG as locomotion team lead before joining Inria where he is currently (having a blast) doing research at the interface between motion control, machine learning and computer vision. Stéphane is a proponent of open source robotics and contributes to projects like Upkie wheeled bipeds, robot_descriptions.py or qpbenchmark.
Feel free to post a comment by e-mail using the form below. Your e-mail address will not be disclosed.