Reproduce “Learning to Optimize in Model Predictive 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
Jacob Sacks, Byron Boots
This work from the University of Washington introduces L2O-MPC, a framework for learning the update rule in sampling-based Model Predictive Control, which consistently achieves higher success rates with substantially fewer samples than traditional MPPI across multiple simulated robotics tasks.
Key insights
- Applying the 'learning to optimize' framework to MPC update rules, rather than just neural network optimization, can improve control performance under sample constraints.
- Feeding a learned optimizer individual trajectory costs alongside structured sampling (e.g., Halton sequences) provides more valuable information than aggregated noisy gradients alone.
- Imitation learning, specifically the DAGGER algorithm, offers an effective method for training these learned optimizers without requiring full differentiability or extensive reinforcement learning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.