TUM AVS

Differentiable Weights-Varying Nonlinear MPC via Gradient-Based Policy Learning (IEEE RA-L)

The paper presents the first differentiable Model Predictive Control (MPC) framework that can vary its cost‑function weights online for constrained nonlinear systems, leveraging gradient‑based policy learning. A lightweight neural network receives real‑time observations—such as reference trajectory curvature and velocity—and outputs MPC weight adjustments at each control step. By back‑propagating a user‑defined loss through a differentiable MPC solver, the authors obtain an end‑to‑end gradient that trains the policy in milliseconds, achieving 38‑times faster convergence and using 27‑times fewer samples than conventional weight‑varying reinforcement learning. Experiments on a high‑fidelity simulation of the full‑scale Delera AV24 race car demonstrate the approach’s potency. The adaptive controller reduces lateral and velocity deviation, cutting path‑tracking error by up to 50 % compared with static‑weight MPC, and matches or exceeds all benchmark algorithms. Moreover, a policy trained on the Monster track transferred zero‑shot to the unseen Laguna Seikka circuit, requiring only two laps of online fine‑tuning to reach performance of a track‑specific controller. These results suggest that fast, sample‑efficient online weight adaptation can eliminate the labor‑intensive tuning traditionally required for MPC, opening the door to more responsive autonomous systems in racing, robotics, and any domain where dynamics shift rapidly.