Risk-Budgeted Control Framework for Balanced Performance and Safety in Autonomous Vehicles

Basic Information

• Paper ID: 2510.10442
• Title: Risk-Budgeted Control Framework for Balanced Performance and Safety in Autonomous Vehicles
• Authors: Pei Yu Chang, Vishnu Renganathan, Qadeer Ahmed (The Ohio State University)
• Classification: eess.SY (Systems and Control), cs.SY (Systems and Control)
• Publication Date: October 12, 2025 (arXiv preprint)
• Paper Link: https://arxiv.org/abs/2510.10442

Abstract

This paper proposes a risk-budgeted monitoring framework combined with a control architecture for safety certification in autonomous driving. The framework employs sliding window monitoring of insufficient barrier residuals or non-zero tail risk to ensure system safety. When safety margins deteriorate, it triggers switching from performance-oriented relaxed control barrier functions (R-CBF) to conservative conditional value-at-risk (CVaR-CBF) constraints. The switching is governed by two real-time triggers: feasibility trigger (FT) and quality trigger (QT) conditions. In the most challenging setting with 5-meter pedestrian detection uncertainty, the framework achieves 94-96% collision-free success rates with pedestrians across 300 trials while maintaining the lowest average crosstrack error (CTE = 3.2-3.6 meters).

Research Background and Motivation

Problem Definition

Safe control of autonomous vehicles under uncertainty is a critical challenge. While traditional control barrier functions (CBF) can guarantee safety, they are only effective when the quadratic program (QP) is feasible and optimal. Existing methods have limitations when facing:

1. Computational Complexity: Uncertainty significantly increases computational demands while controllers must operate at 20-50 Hz frequencies
2. Feasibility Issues: Input constraints and safety constraints may render the internal QP infeasible
3. Performance-Safety Tradeoff: Conservative safety strategies sacrifice performance, while aggressive performance strategies may violate safety requirements

Research Motivation

Existing approaches address feasibility through parameter tuning, hierarchical optimization, or learning-enhanced feasible sets, but often require extensive data collection or cannot directly quantify safety during persistent infeasibility. From a practical perspective, this paper argues that infeasibility does not necessarily imply loss of safety, proposing a risk-budgeted monitoring framework.

Core Contributions

1. Formalized Risk-Budgeted Monitor: Provides windowed safety guarantees within finite time horizons
2. Monitoring-Driven Switching Mechanism: Enables real-time navigation between performance-oriented CBF-QP and conservative CVaR-CBF, limiting tail risk
3. Theoretical and Experimental Validation: Demonstrates that risk budgets allow controlled transient infeasibility while maintaining certified finite-horizon safety

Methodology Details

Task Definition

Consider a nonlinear continuous-time control-affine system: $\dot{x} = f(x) + g(x)u$

where $x \in D \subset \mathbb{R}^n$ and $u \in U \subset \mathbb{R}^m$ denote system state and input respectively. The safety set is defined as $S = \{x : h(x) \geq 0\}$.

Model Architecture

1. Relaxed CBF (R-CBF)

A relaxed version of traditional CBF that handles infeasibility due to actuator limitations by introducing non-negative slack variables $\nu_k$:

$\min_{u_k \in U, \nu_k \geq 0} \frac{1}{2}\|u_k - u_k^{nom}\|^2 + \rho_\nu \nu_k^2$$\text{s.t. } L_f h(x_k) + L_g h(x_k)u_k + \kappa h(x_k) \geq -\nu_k$

2. Stochastic Safety Set and CVaR-CBF

Considering uncertainty in obstacle location and vehicle localization, safety assessment is reformulated as probabilistic risk. Define the loss function: $Z_k^i(u_k) := -r_k^i(u_k)$

where $r_k^i(u_k)$ is the safety residual.

CVaR Constraint: $\text{CVaR}_\epsilon(Z_{k+1}^i) = \inf_{\gamma \in \mathbb{R}} \left\{\gamma + \frac{1}{1-\epsilon}\mathbb{E}[(Z_{k+1}^i - \gamma)^+]\right\}$

Relaxed CVaR-CBF Optimization: $\min_{u_k \in U, \nu_k \in \mathbb{R}} \frac{1}{2}\|u_k - u_k^{nom}\|^2 + \rho_\nu \nu_k^2$$\text{s.t. } \text{CVaR}_\epsilon(Z_{k+1}^i) \leq \nu_k, \quad 0 \leq \nu_k \leq \bar{\nu}$

3. Risk-Budgeted Monitor

Bad Step Detection: Define bad step indicator $b_k = \mathbf{1}\{\nu_k > \bar{\nu}\} \vee \mathbf{1}\{r_{\min,k} < \delta\}$

where $\delta > 0$ is a predefined safety margin.

Sliding Window Monitoring: Within a sliding window of length $W$, at most $M$ bad steps are allowed: $m_k = m_{k-1} + b_k - b_{k-W}$

Window-Level Safety Certification Theorem: $\mu^M(1-\mu^{W-M})\delta \geq (1-\mu^M)\nu$

where $\mu = e^{-\kappa T_s}$.

4. Switching Trigger Mechanism

Feasibility Trigger (FT-C-CBF):