Reproduce “Real-time Model Predictive Control and System Identification Using Differentiable Physics Simulation” 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
Sirui Chen, Keenon Werling, Albert Wu, C. Karen Liu
A dual-threaded framework leverages differentiable physics for real-time model predictive control and adaptive system identification, allowing robots to continuously refine their internal models and control policies post-deployment. This approach achieves superior control performance and rapid parameter identification across various simulated dynamic tasks and on a real robot arm.
Key insights
- Differentiable physics can enable the simultaneous and efficient execution of real-time online system identification and optimal control within a dual-threaded architecture.
- An adaptive online system identification approach, incorporating a confidence score for parametric excitation and weighted averaging, improves robustness against noisy or uninformative data in dynamic environments.
- Integrating actively controlled system identification into the MPC framework allows robots to strategically explore behaviors that maximize parametric excitation, leading to faster and more accurate model adaptation when passive observations are insufficient.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.