When Pedestrians Hesitate: PPO-Based RL Collision Avoidance in Uncertain Scenarios

Pedestrian collision avoidance at unmarked crosswalks remains a critical challenge for Autonomous Vehicles (AVs), particularly when pedestrians exhibit hesitant or unpredictable behaviors. This work introduces a Reinforcement Learning (RL) framework wherein an AV learns to navigate such a scenario by reasoning under behavioral uncertainty. The environment is modeled as a Markov Decision Process (MDP), and the AV is trained using a Proximal Policy Optimization (PPO) algorithm augmented with a Long Short-Term Memory (LSTM) network. The AV selects among four discrete maneuvers: maintain speed, decelerate, dodge left, or dodge right, based on real-time observations. The pedestrian is modeled as a stochastic, contextaware agent with a tunable hesitation probability parameter (phi(h)), modulating its likelihood to stop, reverse, or proceed depending on proximity to the crossing center. A shaped reward function incentivizes safe, timely, and socially compliant actions. Extensive simulations under varying levels of pedestrian hesitation show that the PPO-LSTM agent achieves a collision rate as low as C-R = 5.84% under phi(h) = 0.1. The AV executes evasive maneuvers in 77.03% of decisions and completes episodes in an average of approximate to 3228 steps. These results highlight the agent's capacity to safely and adaptively handle pedestrian indecisiveness without relying on trajectory prediction models or multi-agent RL for AV-pedestrian interaction.