Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Risk Aware Safe Control with Cooperative Sensing for Dynamic Obstacle Avoidance

Pei Yu Chang, Qizhe Xu, Vishnu Renganathan, Qadeer Ahmed

TL;DR

This work addresses safety in autonomous driving under localization and cooperative-sensing uncertainties by combining a Wasserstein barycenter-based fusion of multi-sensor data with a CVaR-informed control barrier function applied within an MPC framework. The WB-CVaR-CBF formulation yields a tractable convex optimization problem that minimizes control deviation while enforcing risk-aware safety through the CVaR surrogate. The authors provide a rigorous on-vehicle demonstration, including ROS2-based software, and show improved safety margins and robustness across scenarios with GPS and V2X disturbances compared to a baseline MPC-CBF. The results indicate that integrating cooperative sensing with distributional risk measures enables practical, real-time safety guarantees for dynamic obstacle avoidance in AVs, with clear pathways for deployment and tuning via the CVaR level parameter.

Abstract

This paper presents the design, development, and on vehicle implementation and validation of a safety critical controller for autonomous driving under sensing and communication uncertainty. Cooperative sensing, fused via a Wasserstein barycenter (WB), is used to optimize the distribution of the dynamic obstacle locations. The Conditional Value at Risk (CVaR) is introduced to form a risk aware control-barrier-function (CBF) framework with the optimized distribution samplings. The proposed WB CVaR CBF safety filter improves control inputs that minimize tail risk while certifying forward invariance of the safe set. A model predictive controller (MPC) performs path tracking, and the safety filter modulates the nominal control inputs to enforce risk aware constraints. We detail the software architecture and integration with vehicle actuation and cooperative sensing. The approach is evaluated on a full-scale autonomous vehicle (AV) in scenarios with measurement noise, communication perturbations, and input disturbances, and is compared against a baseline MPC CBF design. Results demonstrate improved safety margins and robustness, highlighting the practicality of deploying the risk-aware safety filter on an actual AV.

Risk Aware Safe Control with Cooperative Sensing for Dynamic Obstacle Avoidance

TL;DR

This work addresses safety in autonomous driving under localization and cooperative-sensing uncertainties by combining a Wasserstein barycenter-based fusion of multi-sensor data with a CVaR-informed control barrier function applied within an MPC framework. The WB-CVaR-CBF formulation yields a tractable convex optimization problem that minimizes control deviation while enforcing risk-aware safety through the CVaR surrogate. The authors provide a rigorous on-vehicle demonstration, including ROS2-based software, and show improved safety margins and robustness across scenarios with GPS and V2X disturbances compared to a baseline MPC-CBF. The results indicate that integrating cooperative sensing with distributional risk measures enables practical, real-time safety guarantees for dynamic obstacle avoidance in AVs, with clear pathways for deployment and tuning via the CVaR level parameter.

Abstract

This paper presents the design, development, and on vehicle implementation and validation of a safety critical controller for autonomous driving under sensing and communication uncertainty. Cooperative sensing, fused via a Wasserstein barycenter (WB), is used to optimize the distribution of the dynamic obstacle locations. The Conditional Value at Risk (CVaR) is introduced to form a risk aware control-barrier-function (CBF) framework with the optimized distribution samplings. The proposed WB CVaR CBF safety filter improves control inputs that minimize tail risk while certifying forward invariance of the safe set. A model predictive controller (MPC) performs path tracking, and the safety filter modulates the nominal control inputs to enforce risk aware constraints. We detail the software architecture and integration with vehicle actuation and cooperative sensing. The approach is evaluated on a full-scale autonomous vehicle (AV) in scenarios with measurement noise, communication perturbations, and input disturbances, and is compared against a baseline MPC CBF design. Results demonstrate improved safety margins and robustness, highlighting the practicality of deploying the risk-aware safety filter on an actual AV.

Paper Structure

This paper contains 22 sections, 2 theorems, 27 equations, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Preliminaries and Problem formulation
  3. Safe Sets and Control Barrier Functions
  4. Problem Formulation
  5. AV Localization Uncertainty:
  6. Cooperative sensing Measurement Uncertainty:
  7. Cooperative sensing and Wasserstein Barycenter
  8. Example on our three distribution
  9. Wasserstein Barycenter-based Control Barrier Function
  10. Risk-Aware Safety-Critical Control Formulation
  11. WB-CVaR-CBF-OP Formulation
  12. Experimental and Numerical Example: Autonomous Vehicle Control
  13. Experimental Setup
  14. Scenarios and Parameters
  15. Comparative Analysis
  16. ...and 7 more sections

Key Result

Lemma 1

mallasto2017 Let $\{\hat{\mathbf{Z}}_{o,k}^{(s)}\}_{s=1}^{N}$ be a Gaussian measurement with $\hat{\mathbf{Z}}_{o,k}^{(s)} \sim \mathcal{N}(\mathbf{Z}_{o,k}^{(s)}+\mu^{(\text{s})}, \Sigma^{(s)})$, then there exists a unique barycenter $\hat{\mathbf{Z}}_{o,k}^{(\text{WB})} \sim \mathcal{N}({\mathbf{Z

Figures (4)

  • Figure 1: (A) The proposed framework with multiple sensors and wasserstein barycenter with CVaR-CBF (B) The experimental vehicle and the sensors mounted position (C) Overview of the problem statement depicting uncertainty in vehicle position and obstacle position.
  • Figure 2: On-vehicle software architecture and data flow.
  • Figure 3: On-vehicle trajectory with MPC-WB-CVaR-CBF under scenario 2.
  • Figure 4: Trajectories under MPC–WB–CVaR–CBF (proposed) and MPC–CBF–QP (baseline). (a) Scenario 1: both safe; (b) Scenario 2: proposed safe, baseline unsafe; (c) Scenario 3: proposed safe, baseline unsafe.

Theorems & Definitions (2)

  • Lemma 1
  • Lemma 2