Estimation-Aware Reachability

Himanshu Paudel, Javier BorquezarXiv 2026 · arXiv

This paper presents a novel framework for robotic safety that bridges the gap between formal reachability analysis and real-world state estimation. By incorporating a robot's "belief"—its estimate and the associated uncertainty—directly into the safety computation, the authors create safety filters that are both more robust and less conservative than traditional methods.

TL;DR

The authors extend Hamilton-Jacobi (HJ) Reachability to account for noisy state estimates by augmenting the system state with its mean and covariance. By treating measurement noise as an adversarial input and solving the resulting high-dimensional problem with neural PDE solvers, they achieve near-zero collision rates in scenarios where standard safety filters fail.

What Problem Does This Paper Address?

Ensuring safety for autonomous systems often relies on Hamilton-Jacobi (HJ) Reachability. This method calculates a "Backward Reachable Tube" (BRT)—the set of all states from which a system might be forced into a collision, regardless of its control efforts.

However, classical HJ reachability assumes perfect state observability. In reality, robots use sensors to estimate their state, and these estimates are always noisy. As shown in the figure below, applying a "vanilla" safety filter to a noisy estimate is dangerous: the robot might believe it is safe (the blue dot) while its true position (the black dot) has already crossed into the danger zone.

Figure 1: Safety filtering with (a) vanilla HJ and the exact state, (b) vanilla HJ and an estimate ˆp, and (c) estimation-aware HJ under the same estimation error as (b).

Simple fixes, like adding a fixed "safety margin" (padding), often fail because uncertainty isn't static—it evolves based on how the robot moves and where its sensors are most effective.

Methods

The paper introduces Estimation-Aware Reachability, which treats the state estimation process as part of the system dynamics.

  1. Belief-Augmented State: The authors define a new "belief state" $z$ that includes the robot's physical state, its current estimate, and the vectorized covariance matrix (from an EKF or $H_\infty$ filter).
  2. Adversarial Innovation: They treat the measurement innovation (the difference between what the sensors see and what the filter expects) as a bounded adversarial disturbance. This means the safety filter assumes the "worst-case" sensor noise will try to push the robot toward danger.
  3. Uncertainty-Inflated Target Set: The failure set is dynamically expanded based on the current covariance. If the robot is very uncertain, the "danger zone" grows larger in its mind.
  4. DeepReach Solver: Augmenting the state leads to high-dimensional problems (up to 9D in this paper), which are impossible to solve with traditional grid-based methods. The authors use DeepReach, a neural network-based solver that approximates the HJ value function without a grid.
  5. Probabilistic Guarantees: They prove a pointwise probabilistic safety guarantee: if the calculated value function is positive, the true state is safe with a user-defined probability (e.g., $>98\%$).

Datasets

The authors evaluated their framework on three simulated robotic benchmarks:

  • Air3D: A classic pursuer-evader collision avoidance problem.
  • Planar Quadrotor: A 9D system navigating past an obstacle using an $H_\infty$ filter for position estimation.
  • Dubins Car with Position-Dependent Noise: A scenario where sensor noise increases as the car moves further from a beacon, testing the filter's ability to adapt to spatially varying uncertainty.

Key Insights and Ideas

  • Safety is Coupled with Information: The most important "aha moment" is that safety isn't just about physical distance; it's about the quality of information. The filter learns to be more cautious in "dark" areas where sensor noise is high.
  • Innovation as an Adversary: By modeling sensor updates as an adversary in a zero-sum game, the framework naturally handles "jumps" in the estimate that would cause simpler filters to fail.
  • Consistency: The framework is mathematically consistent; if the uncertainty and noise vanish, it perfectly recovers the standard "perfect information" HJ reachability solution.

Key Results

The estimation-aware filter significantly outperformed baselines across all case studies:

  • Air3D: Reduced collision rates from 31.4% (no filter) and 24.3% (nominal filter) down to just 0.1% at high noise levels.
  • Quadrotor: Achieved a 0% collision rate in a scenario where the standard safety filter still collided in 70.8% of rollouts.
  • Dubins Car: The filter demonstrated "intelligence" by engaging much earlier when the car was far from its beacon (high uncertainty) and allowing for tighter maneuvers when close to the beacon (low uncertainty). It reduced intervention costs by 36% compared to a conservative maximum-margin baseline.

Limitations and Future Work

  • Pointwise vs. Trajectory Guarantees: The current mathematical proof provides a guarantee for a single point in time. Extending this to a "trajectory-level" guarantee (safety over a whole time window) is a suggested next step.
  • Worst-Case Conservatism: Treating sensor innovation as a bounded adversary is very safe but can be conservative. Future work could explore stochastic modeling of the innovation to allow for even more agile behavior.
  • Training Time: While DeepReach makes high-dimensional problems solvable, training the neural network still requires significant offline computation (e.g., ~1.8 hours for the Air3D case).

Summary updated

Abstract

Hamilton-Jacobi Reachability computes the states from which a system can be kept safe under worst-case disturbances, but it assumes the state is known exactly, whereas real robots act on noisy estimates whose uncertainty evolves with their motion. We extend Hamilton-Jacobi Reachability to the robot's belief, estimate and uncertainty, treating unexpected measurements as a bounded adversary and enlarging the unsafe set by the current uncertainty, so that certified states are safe with a prescribed probability. The formulation is compatible with existing solvers. In simulation, the resulting safety filters nearly eliminate the collisions of certaintyequivalent HJ and are less conservative than equally safe uncertainty-padding baselines. The implementation code is available at https://github.com/HimanshuPaudel/Estimation-Aware-Reachability
BibTeX
@misc{paudel2026estimation,
  author = {Himanshu Paudel and Javier Borquez},
  title = {{Estimation-Aware Reachability}},
  year = {2026},
  eprint = {2611.23830},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.23830},
}
Figure