Stability- and Safety-Constraint Reinforcement Learning for Pedestrian Avoidance in Occluded Urban Driving

Road traffic accidents continue to be a major global cause of fatalities, disproportionately affecting pedestrians and other vulnerable road users. While deep reinforcement learning has proven effective in handling complex navigation tasks, providing formal stability and safety guarantees during both training and deployment remains a significant challenge. This paper introduces a dual-layer safety-aware framework for pedestrian avoidance in occluded urban driving. During training, a first-order Control Lyapunov–Barrier Function is integrated with Proximal Policy Optimization to promote goal-reaching stability and obstacle avoidance: the analytic Lie derivatives of the Lyapunov and barrier functions are embedded as a modifier in the advantage estimate, providing explicit stability and safety signals that accelerate convergence toward safe, goal-reaching behavior without disrupting the standard policy update. At deployment, a higher-order Control Lyapunov–Barrier Function, realized through a quadratic programming safety filter, acts as a safety shield that projects the nominal acceleration onto the intersection of the second-order Lyapunov and barrier feasibility sets; the barrier function is further extended with relative velocity terms to account for dynamic pedestrian motion. Experiments with a four-wheeled vehicle in the Webots simulator show that the framework reliably reaches the goal, avoids an occluded pedestrian across a range of crossing speeds, and improves task success rates and safety-constraint adherence relative to Proximal Policy Optimization and a conventional higher-order safety filter baseline, particularly during emergency braking maneuvers.