Hamilton-Jacobi (HJ) Reachability identifies the unsafe states of a dynamical system from which it can be forced into failure under worst-case disturbance. However, classical grid-based solvers scale exponentially with state dimension. DeepReach improves scalability by replacing the grid with a neural value function trained on the HJ-Isaacs (HJI) variational inequality (VI), but penalizes both sides of the residual equally. This allows the learned value to lie above the true one and unsafe states to be misclassified as safe. We propose Conservative DeepReach (ConDR) which aims to learn a subsolution of the HJI VI such that the learned unsafe set contains the true one. ConDR enforces the obstacle branch and terminal condition by construction, pushes the HJI residual to a safety margin ε below zero with a one-sided loss, and prevents the degenerate solution of classifying everything as unsafe. We prove that the exact solution of the margin-shifted VI that ConDR trains toward is a subsolution of the original VI so its unsafe set contains the true one and its value lies within ετ of the true value function. We compare against DeepReach and MADR in the established systems of Air3D, Air6D, a 9D multi-vehicle system, and a 13D quadrotor. ConDR reports the lowest rate of unsafe state misclassified and the lowest closed-loop collision rates with at most a small increase in intervention. On Air3D and Air6D, ConDR tolerates model mismatch that scales with ε. At matched false-negative rates, ConDR's unsafe sets are smaller than the baselines showing the safety is not obtained by trivial over-approximation.
Figures & tables
System
Model
IoU ↑
FN ↓
FP ↓
frac (Vθ≤Vtrue)↑
max(Vθ−g)↓
Air3D
DeepReach
0.939±0.008
0.0022±0.0004
0.0059±0.0008
0.659±0.016
+0.018
ConDR
0.929±0.004
0.0009±0.0018
0.0071±0.0006
0.959±0.005
−0.005
MADR
0.952±0.004
0.0266±0.0035
0.0021±0.0001
0.440±0.043
+0.027
ConDR-MADR
0.963±0.002
0.0172±0.0013
0.0020±0.0002
0.740±0.028
−0.005
Air6D
DeepReach
0.950±0.002
0.0123±0.0024
0.0032±0.0003
0.466±0.096
+0.028
ConDR
0.924±0.004
0.0006±0.0001
0.0065±0.0004
0.994±0.001
−0.002
TABLE I: Ground-truth metrics at τ=T on Air3D (5 seeds) and Air6D (3 seeds), mean ± std. FN: unsafe states called safe. FP: safe states called unsafe. frac (Vθ≤Vtrue) : containment. max(Vθ−g) : worst obstacle-branch violation.
Fig. 1: BRT classification at the slices where the baselines miss most. Top two rows: Air3D heading slices. Bottom two rows: Air6D pursuer-position slices with the evader fixed (triangle). Gray: true positives. Red: unsafe states called safe. Blue: safe states called unsafe. Black: true BRT boundary.
DeepReach
ConDR
MADR
ConDR- MADR
Air3D
Violation ↓
0.25
0.01
0.49
0.11
δsub↓
0.018
0.004
0.056
0.096
Shift. vol. ↓
0.116
0.097
0.162
0.152
Air6D
Violation ↓
0.52
0.03
0.50
0.11
δsub↓
0.056
0.033
0.158
0.144
Shift. vol. ↓
0.129
0.094
0.251
0.149
TABLE II: Sampled residual bound ( n=200 K, η=0.01 , β=0.001 ). Violation: fraction of samples violating a subsolution inequality. δsub : bound on the violation holding on a 1−η fraction of the domain with confidence 1−β . Shifted vol.: fraction of the box in {Vθ≤δT} for ConDR and ConDR-MADR (exact obstacle branch) and {Vθ≤δ(T+1)} for DeepReach and MADR (obstacle branch violated up to δ ).
Verified-safe starts
Air3D
Air6D
Method
pursuit
learned
pursuit
learned
No filter
.178
.215
.140
.160
DeepReach
.046 ± .005 /.054
.040 ± .042 /.074
.034 ± .008 /.038
.098 ± .008 /.060
ConDR
.000 ± .000 /.055
.004 ± .005 /.069
.021 ± .005 /.037
.078 ± .010 /.062
MADR
.090 ± .017 /.053
.172 ± .010 /.073
.048 ± .006 /.037
.113 ± .015 /.060
TABLE III: Closed-loop safety filtering: collision rate (mean ± std) / intervention fraction, 400 starts per seed. Top: verified-safe starts on Air3D and Air6D under pure-pursuit and learned adversaries. Bottom: unverified starts on all four systems. Bold: lowest collision rate.
Fig. 3: Air3D and Air6D. Left: collision rate versus mismatch M ; each margin ϵ holds until M≈ϵ . Right: IoU (blue) and certificate violation (red, log) versus ϵ .
Model
Obstacle construction
Ldiff
Ldeg
Lopt
ε
IoU ↑
FN ↓
FP ↓
frac (Vθ≤Vtrue)↑
Subsol. viol. ↓
Collisions ↓
DeepReach
–
–
–
–
–
0.939
0.0022
0.0059
0.659
0.253
0.046 ± 0.005
✓
–
–
–
–
0.963
0.0014
0.0035
0.907
0.234
0.019 ± 0.006
†
✓
✓
–
–
0
0.086
0.0000
1.0000
1.000
0.000
0.348 ± 0.024
✓
✓
✓
–
0
0.919
0.0002
0.0083
0.924
0.170
0.022 ± 0.005
✓
✓
✓
✓
0
0.962
0.0021
0.0035
0.937
0.256
0.013 ± 0.005
✓
✓
✓
–
0.02
0.898
0.0014
0.0106
0.949
0.004
0.004 ± 0.003
TABLE IV: Loss ablation on Air3D (5 seeds). Check marks show active components; rows without Ldiff use the two-sided residual. † Degenerate solution.
Fig. 4: Collision rate (left) and intervention (right) as the filter level c is raised from 0 on verified-safe starts; dotted line at c=ϵ .
Model
FN ≤ 0.05
FN ≤ 0.03
FN ≤ 0.01
Air3D (exact BRT volume 0.0863)
DeepReach
0.0854±0.0008
0.0876±0.0008
0.0900±0.0008
ConDR
0.0827±0.0003
0.0850±0.0003
0.0882±0.0003
MADR
0.0835±0.0001
0.0856±0.0002
0.0886±0.0003
ConDR-MADR
0.0822±0.0001
0.0844±0.0002
0.0884±0.0002
Air6D (exact BRT volume 0.0736)
TABLE V: Volume of {Vθ≤c} after setting c to the smallest value meeting each FN target, mean ± std over seeds (Air3D: 5, Air6D: 3). Bold: smallest in column.
Hamilton-Jacobi (HJ) reachability provides a principled framework for synthesizing safety certificates and robust controllers for safety-critical robotic systems. However, applying reachability analysis to high-dimensional nonlinear systems remains challenging: classical grid-based solvers suffer from the curse of dimensionality, continuous-time neural solvers require accurate spatial value gradients, and reinforcement-learning-based approaches often suffer from weak boundary anchoring and non-stationary adversarial policy optimization. We propose a discrete-time neural reachability framework for control-disturbance-affine systems that learns backward reachable tubes (BRTs) and backward reach-avoid tubes (BRATs) through Bellman-Isaacs value propagation. Our key idea is to combine equation-driven self-supervision with structured policy learning: rather than computing explicit PDE-gradients, we exploit the bang-bang structure of optimal safety interventions to construct approximate teacher actions from gradient-free value probes, converting adversarial actor learning into supervised policy learning. To stabilize long-horizon value propagation, we leverage the learned actor to train the value function backward from the terminal boundary using a windowed temporal curriculum, where each window is used as the boundary condition for the next window. Across benchmark problems up to 80 dimensions, our method learns accurate reachability value functions while improving stability over existing learning-based solvers. We further demonstrate observation-space scalability on F1-tenth racing with over 16,000-dimensional egocentric inputs. The learned safety filter generalizes zero-shot to unseen tracks and transfers to a physical RC car, achieving real-time robust collision avoidance.
Zeyuan Feng, Ali Fuat Sahin, Santiago Thorup +1
Department of Mechanical Engineering, École Polytechnique Fédérale de Lausanne, Switzerland
Hamilton-Jacobi (HJ) reachability provides a mathematically rigorous framework for safe control of dynamical systems, but its practical application is bottlenecked by the computational complexity of solving Hamilton-Jacobi-Isaacs variational inequality PDEs in high dimensions. Physics-informed neural networks (PINNs) have recently emerged as a promising alternative to classical mesh-based solvers, yet their performance is highly sensitive to the choice of collocation sampling. In order to learn accurate safety value functions, existing PINNs-based HJ reachability solvers must rely on complex training pipelines and auxiliary supervision. In this work, we propose STEER2REACH (S2R), a PINNs-based HJ reachability solver that requires minimal modification on top of standard PINNs training. S2R's key contribution is a lightweight, low-overhead adaptive collocation sampling distribution constructed by steering forward trajectories using a combination of the optimal control and disturbance signals induced by the current value function, with injected stochastic exploration noise. We demonstrate that despite its simplicity, S2R achieves competitive--and in some cases improved--performance on safety metrics while reducing relative L2 error across a range of reachability benchmarks compared with SoTA MPC-guided HJ reachability solvers, all without requiring multi-stage training or MPC-based supervision.
Sungje Park, Stephen Tu
with the Ming Hsieh Department of Electrical and Computer Engineering, University of Southern California, Los Angeles, CA, USA.
Hamilton-Jacobi (HJ) reachability provides certified backward reachable tubes (BRTs) for safety-critical robotic systems, but the level-set computation scales exponentially in the state dimension and has historically been performed offline. As autonomy stacks (ADAS/AV, humanoid robotics, and vision-language-action models) increasingly rely on black-box, probabilistic policies, a deterministic safety layer with formal guarantees is an essential complement. We present a CUDA implementation of the canonical WENO5 + TVD-RK3 + Lax-Friedrichs level-set Hamilton-Jacobi-Isaacs solver for 2D, 3D, and 4D dynamics that is fast enough to run inside a robot's control loop. In 32-bit floating point (FP32) on a single workstation GPU (NVIDIA RTX 4000 Ada), it computes the 3D Air3D BRT at 112 Hz, a 4D pursuit-evasion BRT at 61 Hz, and a 2D wind-affected drifter BRT at 74-147 Hz, 5.9-17.5x faster than the JAX reference hj_reachability at matched precision and 1.7-16x faster than a torch.compile reimplementation of the same scheme. In 64-bit floating point (FP64) it agrees with hj_reachability on eight benchmark cases, reproduces Mitchell's Air3D BRT to within 0.04% of the reference cell count, and runs 2.0-4.6x faster. On an embedded NVIDIA Jetson AGX Thor it computes Air3D at 66 Hz (N=41) and the 2D drifter at 54 Hz, 3.8-8.9x faster than hj_reachability on the same device. As a multi-agent application, an in-kernel label-propagation extension recovers the K-agent capture-basin Voronoi tessellation from a single combined HJ solve, independent of fleet size; it runs at 4.0 ms per tick in a closed-loop wildfire-response simulation of six quadcopters under time-varying, non-uniform wind.