Cooptimizing Safety and Performance with a Control-Constrained Formulation

Hao Wang, Adityaya Dhande, Somil Bansal

Abstract

This paper introduces a novel framework for cooptimizing safety and performance in autonomous systems. By transforming difficult state constraints into equivalent control constraints through Hamilton-Jacobi (HJ) reachability analysis, the authors develop a formulation solvable via dynamic programming. This approach provides formal safety guarantees while significantly outperforming existing baselines in computational speed and path optimality.

Introduction

In autonomous robotics, safety and performance are frequently at odds. For instance, a drone must navigate at high speeds to complete a mission (performance) while avoiding obstacles (safety). Mathematically, this is typically framed as a state-constrained optimal control problem. However, these problems are notoriously difficult to solve because standard tools like the Dynamic Programming Principle often fail unless restrictive criteria are met. Existing workarounds, such as safety filtering, Model Predictive Control (MPC), and state-augmentation, often suffer from myopic behavior, local optima, or the "curse of dimensionality."

Methodology: The Control-Constrained Formulation

The authors propose a three-step logic to move constraints from the state space to the control space:

  1. Identify the Safe Set: Using HJ reachability analysis, they calculate a safety value function ($V_s$) that determines if it is physically possible for the robot to remain safe from any given state.
  2. Convert Constraints: Instead of direct state constraints, they define a Set of Safe Controls ($U_s$). This set restricts inputs to only those that keep the robot within the safe zone.
  3. Solve the HJB-PDE: They prove that the value function of this control-constrained problem is a viscosity solution to a specific Hamilton-Jacobi-Bellman Partial Differential Equation (HJB-PDE).
Visualization of the safe control set $U_s(x,t)$ under different safety value conditions. When $V(x,t) > 0$, the system is safe and all controls are available; when $V(x,t) \le 0$, the control set is restricted to maintain safety.

A key technical contribution is proving that the safe control set is lower hemicontinuous, a vital property for ensuring the HJB-PDE has a stable solution. By keeping the state space the same size (avoiding auxiliary variables), the method avoids the exponential increase in complexity typical of state-augmentation methods.

Experimental Evaluation

The method was evaluated using a simulated 2D boat navigation task in a rectangular arena with non-convex obstacles. The objective was to navigate from a start point to a goal while minimizing a quadratic cost (distance) and avoiding obstacles. The controller was tested against 100 random initial states and compared against four baselines: State-constrained DP, MPPI, Safety Filtering, and MPC.

Key Results

The proposed method consistently outperformed all baselines across safety, performance, and efficiency metrics.

Comparative performance of the proposed method against state-of-the-art baselines in a 2D boat simulation.
  • Safety: The method achieved a 100% rollout success rate. The closest state-constrained baseline reached 96%, with failures attributed to numerical inaccuracies in its more complex formulation.
  • Performance: Paths were more efficient; the state-constrained baseline had a 4.48% higher cost, and Safety Filtering had a 23.17% higher cost.
  • Efficiency: Offline computation (solving the PDE) took only 21 minutes, compared to 390 minutes for the state-augmented baseline. Online computation was extremely fast at 0.0015 seconds, making it suitable for real-time robotics.

Conclusion and Future Work

While the grid-based PDE solvers used here are highly efficient for low-dimensional systems, they still face scaling challenges for high-degree-of-freedom robots. Future work will explore using Deep Learning (e.g., "DeepReach") to approximate solutions in higher dimensions and investigating smooth overapproximations to handle non-differentiable points in the safety and value functions.

Summary updated

BibTeX
@article{wang2024cooptimizing,
  author = {Hao Wang and Adityaya Dhande and Somil Bansal},
  title = {{Cooptimizing Safety and Performance with a Control-Constrained Formulation}},
  journal = {IEEE Control. Syst. Lett.},
  year = {2024},
  doi = {10.1109/lcsys.2024.3511429},
}
Figure