Elastic ODYN: Differentiable Optimization for Infeasible Control and Learning in Robotics
Researchers have introduced Elastic ODYN, a novel primal-dual non-interior-point quadratic programming (QP) solver designed to handle infeasible problems in robotics and control. Unlike existing methods that assume feasibility, Elastic ODYN uses smooth squared-$\ell_2$ elastic relaxations to manage conflicting objectives and modeling errors, ensuring stable gradients and convergence even when constraints cannot be met. This framework enables the development of an infeasibility-aware differentiable QP layer and an SQP method, which have been evaluated on various robotics tasks, demonstrating improved robustness and reliability over current state-of-the-art elastic QP solvers. AI