Parameter-Conditioned Reachable Sets for Updating Safety Assurances Online

Javier Borquez, Kensuke Nakamura, Somil Bansal

Abstract

Hamilton-Jacobi (HJ) reachability analysis is a powerful tool for analyzing the safety of autonomous systems. However, the provided safety assurances are often predicated on the assumption that once deployed, the system or its environment does not evolve. Online, however, an autonomous system might experience changes in system dynamics, control authority, external disturbances, and/or the surrounding environment, requiring updated safety assurances. Rather than restarting the safety analysis from scratch, which can be time-consuming and often intractable to perform online, we propose to compute parameter-conditioned reachable sets. Assuming expected system and environment changes can be parameterized, we treat these parameters as virtual states in the system and leverage recent advances in high-dimensional reachability analysis to solve the corresponding reachability problem offline. This results in a family of reachable sets that is parameterized by the environment and system factors. Online, as these factors change, the system can simply query the corresponding safety function from this family to ensure system safety, enabling a real-time update of the safety assurances. Through various simulation studies, we demonstrate the capability of our approach in maintaining system safety despite the system and environment evolution.

Introduction

In safety-critical robotics—such as autonomous drones or self-driving cars—engineers use Hamilton-Jacobi (HJ) Reachability Analysis to calculate "safe zones." This mathematical framework identifies states to avoid to ensure safety even under worst-case scenarios. However, these calculations are typically performed offline and assume static environments. If a system encounters unexpected conditions, such as motor degradation or shifting winds, its pre-calculated safe zone becomes invalid. Recomputing these zones from scratch is mathematically intense and too slow for real-time operation. This paper addresses the challenge of online safety adaptation by pre-computing reachable sets that are conditioned on changing parameters.

Methodology

The proposed approach transforms the online re-computation problem into a real-time sampling problem through three key steps:

  1. Parameters as "Virtual States": Environmental factors (like wind speed) or system limits (like maximum steering angle) are treated as additional dimensions in the system's state space. By modeling these parameters as states with zero dynamics ($\dot{\beta} = 0$), the authors can solve for a single high-dimensional safety set that encompasses all possible parameter variations.
  2. DeepReach for High Dimensions: Traditional reachability methods suffer from the "curse of dimensionality," typically failing beyond 5 or 6 dimensions. The authors utilize DeepReach, a framework that uses a neural network (specifically a Sinusoidal Representation Network or SIREN) to solve the underlying partial differential equations (PDEs).
  3. Offline Training, Online Query: The neural network is trained offline to learn a continuous "family" of safety sets. At runtime, when the robot senses a change, it feeds the new parameter value into the network, which instantly outputs the updated safety boundary and control strategy.
Parameter-conditioned reachable sets for varying maneuverability limits ($\bar{\omega}_a(\beta_u)$). The safe set (teal) adapts continuously as the parameter changes.

Experimental Evaluation

The method was evaluated on four increasingly complex simulation benchmarks:

  • Air3D (4D): A collision avoidance scenario where an evader experiences a sudden change in maneuverability.
  • Rocket Landing (7D): A 6-DOF rocket landing on a platform drifting due to ocean currents.
  • Human-Robot Interaction (5D): Navigation around a human where the robot's confidence in human intent shifts based on observation.
  • Three-Vehicle Collision Avoidance (12D): A high-dimensional test involving three agents with varying collision radii.
Simulation of the Air3D scenario. The adaptive controller successfully maintains a safe distance $d(e,p)$ from the evader despite dynamic changes in maneuverability.

Key Results and Discussion

  • Real-Time Adaptation: In the Air3D benchmark, the parameter-conditioned controller recognized sudden engine failures immediately, successfully avoiding collisions that caused static controllers to fail.
  • Scalability: The authors successfully computed a safety set for a 12-dimensional problem. This is a significant computational milestone, as traditional grid-based tools cannot handle such high dimensionality.
  • Efficiency: While offline training for a 4D problem takes approximately 75 minutes on a consumer GPU, the online "query" time is in the millisecond range, making it ideal for real-time deployment.
  • Continuity: Unlike discrete "multi-mode" systems that switch between a few pre-set safety levels, this method is continuous, allowing for smoother and more precise safety transitions across the entire parameter range.

Limitations: Because the safety sets are learned by a neural network, they are approximations. While highly effective in practice, they currently lack the formal mathematical guarantees provided by traditional grid-based methods. Additionally, the system's safety depends on the accuracy of the robot's online parameter estimation.

Conclusion

By treating uncertain parameters as virtual states and leveraging neural PDE solvers, this work enables autonomous systems to maintain safety assurances in dynamic, evolving environments. Future work will focus on developing formal verification for these neural approximations and validating the approach on physical hardware, such as drones in variable wind conditions.

Summary updated

Abstract

Hamilton-Jacobi (HJ) reachability analysis is a powerful tool for analyzing the safety of autonomous systems. However, the provided safety assurances are often predicated on the assumption that once deployed, the system or its environment does not evolve. Online, however, an autonomous system might experience changes in system dynamics, control authority, external disturbances, and/or the surrounding environment, requiring updated safety assurances. Rather than restarting the safety analysis from scratch, which can be time-consuming and often intractable to perform online, we propose to compute \textit{parameter-conditioned} reachable sets. Assuming expected system and environment changes can be parameterized, we treat these parameters as virtual states in the system and leverage recent advances in high-dimensional reachability analysis to solve the corresponding reachability problem offline. This results in a family of reachable sets that is parameterized by the environment and system factors. Online, as these factors change, the system can simply query the corresponding safety function from this family to ensure system safety, enabling a real-time update of the safety assurances. Through various simulation studies, we demonstrate the capability of our approach in maintaining system safety despite the system and environment evolution.
BibTeX
@inproceedings{borquez2023parameter,
  author = {Javier Borquez and Kensuke Nakamura and Somil Bansal},
  title = {{Parameter-Conditioned Reachable Sets for Updating Safety Assurances Online}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2023},
  url = {https://openreview.net/forum?id=qQNXofY4gu},
}
Figure