Brief

Researchers have developed a novel deep unfolding technique to accelerate solvers for large-scale conic optimization problems, particularly semidefinite programs (SDPs) common in robotics. This method addresses memory and numerical stability issues encountered when backpropagating through full-update conic solvers. The new approach utilizes an implicit differentiation rule for memory efficiency and a robust backward rule for PSD cone projections, enabling the learning of lightweight hyperparameter policies and warm-starts. Evaluations show significant speedups, with learned policies outperforming state-of-the-art solvers and achieving up to a 50x speedup on various problems, including those solved via sequential convex programming. AI