A Human-Like Pedestrian Model for Automated Driving Simulations

Developing safe automated vehicles requires high-fidelity models of how pedestrians interact with traffic. Existing simulation models are often too simplistic, relying on rigid rules or failing to capture the adaptive, human-like nature of crossing decisions. The authors seek to create a unified, simulator-ready pedestrian model that reproduces documented human behaviors across diverse, complex traffic conditions.

The researchers formulate pedestrian crossing as a partially observable Markov decision process (POMDP). This framework accounts for the fact that pedestrians have limited, noisy information about approaching vehicles and must make decisions under cognitive and motor constraints. The model is trained using deep reinforcement learning (PPO) in the CARLA simulator, with domain randomization to ensure it encounters varied road geometries and traffic flows. The agent is rewarded for balancing safety, efficiency, and effort, allowing complex behaviors to emerge naturally rather than through manual scripting.

The model successfully reproduces ten key empirical crossing phenomena, including hesitation before entering the road, increased impatience during long waits, and adaptive strategies like rolling crossings in dense traffic. It also demonstrates realistic responses to danger, such as aborting a crossing or performing evasive movements (running or stepping back) when vehicle time-to-arrival is low. In closed-loop simulations, the model significantly outperforms the default CARLA AI walker, achieving a lower collision rate. Furthermore, the model shows strong generalization, retaining its competence when transferred to unseen maps and adapting its behavioral style to different local traffic norms through lightweight fine-tuning.

This work provides a blueprint for creating more realistic pedestrian agents in driving simulators. By moving beyond simple rule-based walkers, this approach allows developers to stress-test automated vehicles against a wider range of human-like behaviors, such as hesitation and tentative entry. This is critical for ensuring that automated systems can safely navigate complex, real-world urban environments where pedestrian behavior is highly adaptive and context-dependent.