Safety Foundations

Reachability, control, planning, verification, and learning

Lifecycle safety requires rigorous tools for reasoning about risk, recoverability, uncertainty, and intervention. In this theme, we develop theoretical and computational foundations for safety-critical autonomy, including reachability analysis, Hamilton-Jacobi methods, control-theoretic safety, planning, verification, and learning-based safety representations.

Representative Projects

DeepReach: A Framework for High-Dimensional Reachability Analysis

DeepReach learns neural approximations of high-dimensional Hamilton-Jacobi reachability value functions, enabling safety analysis and controller synthesis where grid-based methods do not scale.

Paper 1Paper 2Code

Verification of Neural Reachable Tubes

Scenario optimization and conformal prediction turn learned reachable tubes into formal safety assurances, providing finite-sample guarantees for neural reachability methods.

Paper

Safety Under State Uncertainty

Robust verification under state uncertainty uses Hamilton-Jacobi reachability to reason about controller safety when the system state is only partially known through noisy perception or localization.

PaperWebsite

Reachability for Black-Box Dynamical Systems

This project computes reachable sets and safe controllers when only a black-box dynamics model is available, estimating the Hamiltonian from samples and integrating with existing reachability solvers.

PaperCode

DualGuard MPPI

DualGuard MPPI combines sampling-based model predictive control with Hamilton-Jacobi reachability so a performant controller can be guarded by principled safety reasoning.

Paper

Physics-Informed Learning for Co-Optimizing Safety and Performance

Physics-informed learning embeds dynamics and control constraints into neural optimization, helping autonomous systems co-optimize task performance and safety for nonlinear control problems.

Paper

Safe Spacecraft Docking

Neural backward reach-avoid tubes with MPC supervision scale reachability-style safety analysis to high-dimensional spacecraft docking, learning safety certificates and recovery guidance for safe approach.

Paper

Reachability for Hybrid Systems

Hamilton-Jacobi reachability for hybrid systems handles continuous dynamics with controlled and forced mode transitions, enabling safety analysis for systems such as legged robots with contact switches.

PaperWebsiteCode