Reliable Conditioned Hamilton-Jacobi Reachability via Subsolution Learning on Exact Local Windows

Seunghwan Jang, Wei XiaoarXiv 2026 · arXiv

TL;DR

This paper introduces a novel approach to Hamilton-Jacobi (HJ) reachability that allows robots to navigate new, complex environments without the need for time-consuming re-computation. By framing the problem as "subsolution learning" within a local window, the authors create a neural network that is conservative by design—ensuring that if the model claims a state is safe, it truly is, even in rare or crowded obstacle configurations.

What Problem Does This Paper Address?

In autonomous systems, Hamilton-Jacobi (HJ) reachability is the gold standard for safety. It provides the "maximal safe set"—the entire region from which a robot can guaranteed-avoid collisions—and a corresponding safety controller.

However, HJ reachability suffers from two major bottlenecks:

  1. The Curse of Dimensionality: Traditional grid-based solvers scale exponentially, limiting them to systems with fewer than six state variables.
  2. The Re-solve Problem: If the environment changes (e.g., a new obstacle appears), the entire safety value function must be re-calculated. While neural solvers like DeepReach can handle higher dimensions, they take minutes or hours to train for a single specific scene.

Previous attempts to learn a "conditioned" model (one network for all scenes) often fail because neural networks make errors in both directions. A "false-unsafe" error makes the robot timid, but a "false-safe" error leads to a crash. In safety-critical robotics, we need errors to be conservative—falling only on the side of caution.

Methods

The authors propose a "window-conditioned" reachability model built on three pillars:

  • Exact Local Windows: The authors prove that by capping the obstacle margin (the distance to the nearest obstacle) at a specific threshold, the safety value becomes provably local. Obstacles beyond a certain radius simply cannot affect the robot's immediate safety. This allows the model to look only at a small "body-frame window," enabling it to train on small scenes and transfer to massive ones instantly.
  • Subsolution Learning: Instead of training the network to match the safety value exactly (which leads to unpredictable errors), they enforce a subsolution constraint. The network architecture is designed to never exceed the obstacle margin, and the loss function penalizes violations of the HJ inequality. This ensures the learned safe set grows outward from a conservative core toward the true safe set.
  • Counterexample Mining: To ensure the model handles rare, dangerous configurations (like tight "dead-ends"), the authors use adversarial mining. They search the finite-dimensional space of possible window configurations to find and retrain on the most difficult cases, rather than relying on random scene sampling.

Datasets

The method was evaluated on two primary robotic benchmarks:

  1. 3D Dubins Car: Tested across 12 distinct obstacle settings, including "dead-ends," cluttered environments, and scenes 6.5x larger than the training area.
  2. 10D Quadrotor: A high-dimensional challenge involving flight through sparse and dense spherical obstacles in large rooms.
  3. Baselines: The model was compared against six state-of-the-art conditioned or compositional methods, including DeepReach, ORN-CBF, and certFNO.

Key Insights and Ideas

  • Equality is Dangerous: The authors demonstrate that standard "equality-residual" training (trying to fit the HJ equation exactly) is inherently biased toward optimism. In safety, optimism kills. Subsolution training is the necessary mathematical pivot for reliability.
  • Locality is a Choice, Not a Compromise: By capping the margin $\delta$, locality becomes a mathematical property rather than a heuristic. This is what allows a model trained on a $2m \times 2m$ patch to work perfectly in a $24m \times 24m$ room.
  • Conservative Merging: When a window contains more obstacles than the model can process (e.g., more than six), the model merges them into larger covering disks. Because the safety value is monotonic, this merging is guaranteed to be conservative—it might lose some safe space, but it will never create a false sense of safety.

Key Results

  • Zero Collisions: In the Dubins car experiments, the proposed method was the only one to achieve a 0% collision rate across all nine standard settings. Baselines had collision rates as high as 14.6%.
  • High Recovery: Despite being conservative, the model "recovered" (identified) over 98% of the true safe set in most scenarios, comparable to models that took hours to train on a single scene.
  • High-Dimensional Success: On the 10D quadrotor, the model maintained significantly lower collision rates (0.004% to 0.16%) than DeepReach (up to 13.2%) and other baselines.
  • Real-Time Performance: The model prepares a new scene in milliseconds and answers safety queries in ~1ms on a CPU, making it fast enough for 100 Hz safety filters.

Limitations and Future Work

While the results are robust, the authors note a few limitations:

  • No Formal Soundness: The safety is verified empirically against grid tolerances; there is no formal proof that the neural network will always satisfy the HJ inequality everywhere.
  • Disk-Only Obstacles: Currently, the model represents obstacles as disks or spheres. While other shapes can be "covered" by disks, it introduces extra conservatism.
  • Avoidance Only: The current formulation focuses on avoiding obstacles. Future work is needed to extend this to "reach-avoid" problems, where the robot must also reach a specific goal.

Reference: Jang, S., & Xiao, W. (2026). Reliable Conditioned Hamilton-Jacobi Reachability via Subsolution Learning on Exact Local Windows. arXiv:2611.19088v1.

Summary updated

Abstract

Hamilton-Jacobi (HJ) reachability computes the maximal safe set and a safety filter for a given obstacle configuration, but it must be re-solved whenever the obstacles change, which a deployed robot cannot afford. Learning a single value function conditioned on the obstacle set removes the re-solve, yet it sharpens the central weakness of learned reachability: one network must now cover a whole family of safety problems, and its approximation error carries no preferred sign, so rare obstacle configurations silently produce false-safe states. To address these challenges, we present a conditioned reachability model designed so that the remaining error falls on the conservative side. Specifically, our method caps the obstacle margin, which makes the value function provably local, in the sense that obstacles farther than a known threshold cannot affect it, so the model reads only a body-frame window, trains on small scenes, and transfers to scenes of any size with millisecond per-scene preparation. Moreover, the output structure builds the value bound into the model. By training with counterexample mining over the finite-dimensional window configuration space, we maximize the claimed safe volume subject to the HJ inequality, so the learned set grows toward the true safe set from the conservative side. Across all tested obstacle settings, ours is the only method that combines recovery comparable to per-scene DeepReach training with near-zero rollout collisions. We also extend the method to a 10D quadrotor, again with markedly lower collision rates than the baselines.
BibTeX
@misc{jang2026reliable,
  author = {Seunghwan Jang and Wei Xiao},
  title = {{Reliable Conditioned Hamilton-Jacobi Reachability via Subsolution Learning on Exact Local Windows}},
  year = {2026},
  eprint = {2611.19088},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.19088},
}
Figure