ConDR: Conservative Reachability through Neural Hamilton-Jacobi Subsolutions

This paper presents ConDR (Conservative DeepReach), a method for computing conservative reachability sets in dynamical systems using neural networks. Unlike previous approaches that may misclassify unsafe states as safe, ConDR enforces the learned value function to be a subsolution of the Hamilton-Jacobi-Isaacs variational inequality, ensuring safety guarantees by construction. The authors demonstrate ConDR on various benchmarks including Air3D, Air6D, multi-vehicle systems, and quadrotors, showing improved safety margins and robustness to model mismatch compared to baselines like DeepReach and MADR.