Learning What Matters: Task Tailored Dynamics Models through Differentiable MPC
Jan Węgrzynowski, Piotr Kicki, Grzegorz Czechmanowski, Walas, Krzysztof, Tadeusz
Abstract
In model-based control, dynamics models are typically trained by minimizing open-loop prediction errors uniformly across all states. However, due to finite model capacity, this misallocates representational power, as not all prediction errors impact the downstream closed-loop performance equally. In this extended abstract, we propose a task-aware training methodology for a prediction model used in the context of Model Predictive Control (MPC). By extracting analytical sensitivities via differentiable MPC, we construct a loss function that weights multi-step dynamics model prediction errors based on their impact on the closed-loop task cost. Experimental results on a simulated 7DoF manipulator demonstrate that our sensitivity-weighted loss significantly improves closed-loop tracking performance compared to standard Mean Squared Error (MSE) or variance-based state standardization.