Learning What Matters: Task Tailored Dynamics Models through Differentiable MPC.

Jan Węgrzynowski, Grzegorz Czechmanowski, Piotr Kicki, Krzysztof Walas

Keywords: Model Predictive Control, Dynamics Learning, Differentiable Optimization, System Identification

TL;DR: We propose a task-aware training objective for MPC dynamics models that uses differentiable MPC to weight prediction errors based on their approximated impact on closed-loop performance, significantly improving downstream control.

Abstract:

In model-based control, dynamics models are typically trained by minimizing open-loop prediction errors uniformly across all states and time steps. However, due to finite model capacity, this approach often misallocates representational power, as not all prediction errors impact the downstream closed-loop task equally.
In this paper, we propose a task-aware training methodology that weights multi-step prediction errors based on their analytical sensitivity approximation to the closed-loop task cost, extracted via differentiable MPC. Experimental results demonstrate that fine-tuning dynamics models with our sensitivity-weighted loss significantly improves closed-loop tracking performance compared to standard Mean Squared Error (MSE) or variance-based state standardization.

Submission Number: 18