Reproduce “DeepReach: A Deep Learning Approach to High-Dimensional Reachability” 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

Somil Bansal, Claire Tomlin

DeepReach introduces a deep learning framework for Hamilton-Jacobi reachability analysis, employing a self-supervised approach with sinusoidal neural networks to compute accurate safety-guaranteeing sets and controllers for high-dimensional nonlinear systems. This method successfully addresses problems up to 10 dimensions, previously intractable for traditional verification techniques.

Key insights

  • The HJI variational inequality can effectively serve as a self-supervision loss function for DNN training, enabling the solution of high-dimensional PDEs without prior knowledge of their solutions.
  • Sinusoidal activation functions are superior for neural PDE solvers, as they accurately represent both a signal and its derivatives, which is crucial for Hamiltonian computations and robust controller synthesis in reachability analysis.
  • The deep learning approach's computational complexity scales with the 'signal complexity' of the value function rather than exponentially with state space dimensionality, enabling analysis of previously intractable high-dimensional systems.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.