Learning-Guided Risk-Adaptive MPC with Control Barrier Functions for Safe Navigation in Uncertain Dynamic Environments

Zhirui Sun1,2, Siyuan Wang1, Yuhan Pang1, Bi-Ke Zhu1 and Jiankun Wang1,2

 1Shenzhen Key Laboratory of Robotics Perception and Intelligence, Department of Electronic and Electrical Engineering, Southern University of Science and Technology, Shenzhen, China.

2Jiaxing Research Institute, Southern University of Science and Technology, Jiaxing, China.

Abstract: Safe navigation among dynamic obstacles is challenging due to complex interactions and unavoidable sensing and prediction uncertainty. We propose a learning-guided, risk-adaptive control framework, Risk-Adaptive Model Predictive Control with Control Barrier Functions (RA-MPC-CBF). A policy trained with Proximal Policy Optimization (PPO) adaptively adjusts a risk level β online, then maps it to a risk-sensitive safety margin for the CBF constraints. During training, the policy interacts with an uncertainty-aware CVaR-CBF controller that provides real-time feedback on solution feasibility and safety performance. At deployment, β is mapped to a risk-adaptive margin within an MPC-CBF, which executes the first feasible plan from a prioritized candidate set. Feasibility-oriented reward design and quantile-based inference mechanisms further enhance training stability and deployment robustness. Systematic evaluations in randomized dynamic environments show that RA-MPC-CBF consistently improves safety while maintaining efficiency across varying obstacle densities and noise levels.