FAME-QP: A GPU-Accelerated Quadratic-Programming Solver for Real-Time Model Predictive Control

FAME-QP. A GPU-Accelerated Quadratic-Programming Solver for Real-Time Model Predictive Control. By Santosh Mohan Rajkumar and two others, from the Ohio State University. Published on the tenth of November.

Linear model predictive control gives rise to dense quadratic programs, and those programs have to be solved for the controller to act. The title frames the target as real-time control, where what counts is how long each solve takes. The comparison that matters here is against other solvers that already run on the GPU, and the measure is latency.

FAME-QP is a GPU-accelerated solver for exactly these dense programs. At its core it combines two things. The first is operator representations that depend on the constraints. The second is fused iterations, in single-kernel and two-kernel forms, which are replayed through CUDA graphs.

Around that core, several choices are made automatically rather than by hand. The execution configuration is selected through device calibration together with timing specific to the problem being solved. To choose the initial penalty, the solver probes a cached bank of operators. And when the problem matrices change, the operators are updated lazily.

The evaluation comes in two parts. Numerical benchmarks test the solver on its own. Simulated control experiments then put it to work, steering a humanoid and a quadrotor. Across both, FAME-QP shows reduced latency relative to competing GPU solvers, while maintaining solution accuracy and control performance.

Put together, FAME-QP is a solver that tunes itself twice: once to the device it runs on, and once to the problem it is handed. What it reports in return is lower latency than the competing GPU solvers, with accuracy and control performance kept where they were.