Reproduce “Differentiable MPC for End-to-end Planning and Control” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Brandon Amos, Ivan Dario Jimenez Rodriguez, Jacob Sacks, Byron Boots, J. Zico Kolter
Amos et al. present an efficient method for differentiating through Model Predictive Control (MPC), allowing its parameters to be learned end-to-end from task-specific losses in continuous control settings. The framework achieves greater data efficiency for structured policies in imitation learning compared to neural networks and outperforms traditional system identification when expert dynamics are not perfectly modeled.
Key insights
- Implicit differentiation through the KKT conditions of the final convex approximation at an MPC solver's fixed point enables efficient and accurate gradient computation.
- The backward pass computation for differentiable MPC can be formulated as solving a modified LQR problem, reusing forward pass computations for significant efficiency gains.
- Optimizing dynamics and cost parameters directly against a task-specific loss can yield superior policy performance compared to optimizing for dynamics accuracy alone, especially when the expert model is complex or non-realizable.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.