Reproduce “Parallel Differentiable Reachability for Learning and Planning with Certified Neural Dynamics and Controllers” 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
Keyi Shen, Glen Chou
Researchers at the Georgia Institute of Technology developed DiffReach, a framework that unifies Taylor model-based reachability with CROWN-style neural network verification. This enables certified training of neural dynamics and controllers, and facilitates robust, reachability-aware online planning for robotic systems operating under uncertainty.
Key insights
- Unifying Taylor models for continuous dynamics with CROWN-style verification for neural networks within a single, differentiable JAX framework allows for efficient, sound, and GPU-accelerated reachability analysis.
- Exposing differentiable reachable-set objectives enables direct integration of formal safety guarantees into the training of neural dynamics and controllers, leading to 'verified-by-construction' models.
- Reachable-set information can be effectively incorporated into Model Predictive Control for online planning, enabling robust decision-making that explicitly accounts for bounded uncertainty and avoids unsafe regions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.