Parameter-Conditioned Reachable Sets for Updating Safety Assurances Online

Javier Borquez, Kensuke Nakamura, Somil Bansal

TL;DR

This paper introduces a method to update robot safety guarantees in real-time as environments or system dynamics change (e.g., a sudden motor failure or shifting wind). By treating uncertain parameters as "virtual states" and solving a high-dimensional reachability problem offline using deep learning, the authors enable robots to instantly query an updated safety controller during operation.

What Problem Does This Paper Address?

In safety-critical robotics—like autonomous drones or self-driving cars—engineers use Hamilton-Jacobi (HJ) Reachability Analysis to calculate "safe zones." This math tells a robot exactly which states to avoid to stay safe, even in the worst-case scenario.

The catch? These calculations are usually done before the robot is deployed and assume the world stays the same. If a drone encounters stronger-than-expected winds or a component fails, its pre-calculated "safe zone" becomes invalid. Recomputing these zones from scratch is mathematically intense and takes far too long to do while the robot is mid-flight. This paper tackles the challenge of online safety adaptation: how can a robot update its safety assurances the moment its situation changes?

Methods

The authors propose Parameter-Conditioned Reachable Sets. Their approach breaks down into three main steps:

  1. Parameters as "Virtual States": Instead of treating environmental factors (like wind speed) or system limits (like maximum steering angle) as constants, they treat them as additional dimensions in the system's state space. If a robot has 3 physical states and 2 uncertain parameters, the authors treat it as a 5D system where the parameters simply don't change over time ($\dot{\beta} = 0$).
  2. DeepReach for High Dimensions: Traditional reachability methods scale exponentially with dimensions (the "curse of dimensionality"), making 5D or 6D problems nearly impossible. The authors leverage DeepReach, a recent 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 "family" of safety sets across the entire range of possible parameters. At runtime, when the robot senses a change (e.g., "my max thrust just dropped by 20%"), it simply feeds the new parameter value into the neural network. The network instantly outputs the updated safety boundary and the corresponding control strategy.

Datasets

As this is a control theory and robotics paper, it does not use static image or text datasets. Instead, it evaluates the method on four increasingly complex simulation benchmarks:

  • Air3D (4D): A classic collision avoidance scenario between two vehicles, where the "evader" vehicle experiences a sudden change in maneuverability (simulating an engine failure).
  • Rocket Landing (7D): A 6-degree-of-freedom rocket attempting to land on a platform that is drifting due to ocean currents.
  • Human-Robot Interaction (5D): A robot navigating around a human, where the robot's "confidence" in the human's intent changes based on observed behavior.
  • Three-Vehicle Collision Avoidance (12D): A high-dimensional test involving three agents with varying collision radii.

Key Insights and Ideas

  • The "Virtual State" Hack: By moving parameters into the state vector, the authors transform a re-computation problem into a sampling problem. We don't need to solve a new PDE online; we just need to look up a different slice of a high-dimensional solution we already found.
  • Continuous vs. Discrete: Unlike "multi-mode" safety systems that switch between a few pre-set safety levels (e.g., "Low Wind" vs. "High Wind"), this method is continuous. It can handle any value within the parameter range, allowing for much smoother and more precise safety transitions.
  • Overcoming the Curse of Dimensionality: The paper demonstrates that neural PDE solvers can handle up to 12 dimensions effectively. This is a significant leap over traditional grid-based methods which typically top out at 5 or 6 dimensions.

Key Results

  • Real-Time Adaptation: In the Air3D example, a robot using a "nominal" (non-adaptive) controller crashed after an engine failure because it didn't realize its safe zone had shrunk. The parameter-conditioned controller recognized the danger immediately and successfully avoided the collision.
  • Precision Landing: In the Rocket Landing simulation, the adaptive controller successfully accounted for a landing pad shifting by meters, whereas the standard controller missed the pad entirely.
  • Scaling to 12D: The authors successfully computed a safety set for a 12-dimensional problem (3 vehicles + 3 parameters). This is a massive computational feat that would be impossible with standard tools.
  • Efficiency: Training the neural network for the 4D problem took about 75 minutes on a single consumer GPU (RTX 3090Ti). Once trained, the "query" time is in the millisecond range.

Limitations and Future Work

  • Approximation vs. Guarantee: Because the safety sets are learned by a neural network, they are approximations. Unlike traditional methods, there isn't yet a formal mathematical proof that the neural network's output is 100% accurate, though the authors point to emerging research that might solve this.
  • Parameter Estimation: The system assumes the robot can accurately sense or estimate the changing parameters (like wind speed or failure state). If the sensor data is noisy or wrong, the safety assurance could fail.
  • Hardware Validation: All results in this paper are from simulations. The next logical step is to demonstrate this running on physical hardware, such as a real drone in a wind tunnel.

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 timeconsuming 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.
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},
}

Loading related content...

Figure