FAME-QP preprint reports lower GPU solver latency in simulated robot control

Researchers report that FAME-QP, a GPU-based optimization solver, reduced latency relative to competing GPU solvers in numerical benchmarks and simulated humanoid and quadrotor control experiments. The arXiv preprint says it maintained solution accuracy and control performance. The solver handles quadratic optimization problems used in model predictive control, which calculates control actions from a model. It combines GPU operations into single- or two-kernel iterations replayed through CUDA graphs and selects execution settings through device calibration and problem-specific timing. It also delays operator updates until needed when problem matrices change. The reported control experiments used simulation.