Safety filters stop re-solving per scene: a conditioned reachability model pushes error to the conservative side and collides zero times across nine obstacle settings
Synopsis
Every time a deployed robot's obstacle layout changes, Hamilton–Jacobi reachability must be re-solved, yet learning a single value function conditioned on the obstacle configuration can silently produce false-safe states. A team at Nanyang Technological University caps the obstacle margin so the value depends only on a local window in the robot's body frame, builds the value bound into the output structure, and maximizes the claimed safe volume subject to the HJ inequality. On a Dubins car, the model is the only one among nine obstacle settings within window capacity with no false-safe state within grid tolerance and zero closed-loop rollout collisions, while keeping recovery comparable to per-scene DeepReach training.
Interpretation
One trained model can serve scenes of any size and obstacle count, with millisecond preparation for a new scene instead of a re-solve or retraining. Prior routes either re-solved with grid solvers whose cost grows exponentially with state dimension, or used self-supervised neural solvers such as DeepReach that spend minutes to hours training on each obstacle configuration. Trained only on small scenes, the Dubins model reaches recovery 0.980 in the large setting with 36 disks; preparing a new scene takes 0.03–10 ms against 53–248 ms for grid re-solving.
The model's remaining error is designed to fall on the conservative side, so states it claims safe are not truly unsafe. A conditioned model must cover a whole family of safety problems, its approximation error has no preferred sign, and rare obstacle configurations silently produce false-safe states; equality-residual training is argued to be optimistically biased. Across the nine Dubins settings within window capacity, the model has no false-safe state within grid tolerance and no collision, while all six conditioned or compositional baselines produce false-safe states or collisions somewhere, with collision rates up to 14.6%.
Training mines counterexamples over the finite-dimensional window configuration space, reaching regions that random scene sampling leaves thin. Existing pipelines still draw training scenes at random, and false-safe errors concentrate in crowded obstacle clusters and long horizons that data rarely covers. The violation rate is measured on 2 million held-out points of the window configuration space never used in training, but the specific values for mining on, mining off, and matched collocation count are absent from the text read here, so only the measurement setting itself is reported.
The method extends to a 10D quadrotor with the lowest collision rate in all three obstacle settings. At high dimension grid ground truth is unavailable, so baselines that need true values drop out of the comparison. On the quadrotor in sparse, dense and nine-times-larger rooms, rollout collisions are 0.004–0.16%, against 0.05–0.60% for PCR+ExactBC and 0.10–13.2% for per-scene DeepReach.
Perspective
The demonstrated scope is avoid problems on a 3D Dubins car and a 10D quadrotor, with obstacles as disks in the Dubins scenes and spheres in the quadrotor scenes, and a window capacity of six obstacles; crowded settings beyond capacity are handled by merging. The authors propose formally verifying the HJ inequality, handling general obstacle shapes, and extending the method to reach–avoid problems.
The authors assert no formal soundness: the value bound is built into the output structure, while the HJ inequality is a trained target verified only empirically within grid tolerance. In the three crowded settings, merging obstacles costs recovery, and all failures of the model occur there; false-safe states lie outside the trained configuration range and are flagged by a runtime range check. The method is restricted to avoid problems, since reach and reach–avoid specifications need global value information a window cannot carry.
