Pedestrian Trajectory Prediction With Reinforcement Learning From a Human-Designed Reward Model

Accurately predicting pedestrian trajectories in safety-critical scenarios is crucial for autonomous vehicles. In such contexts, human behaviors, i.e., pedestrians’ responses when noticing and attempting to avoid dangerous vehicles, play a decisive role in determining safe outcomes. Most current pedestrian trajectory prediction models are trained on data collected under low-risk scenarios. Consequently, although they perform well in routine domains, their generalization to high-risk interactions involving human reactions remains limited due to both data bias and models’ structural inadequacies. This article proposes a fine-tuned transformer trajectory prediction model with reinforcement learning from a human-designed reward model under proximal policy optimization. This design effectively mitigates the performance degradation caused by data bias and enhances prediction capability for rare behaviors. To support model training and evaluation, we constructed a pedestrian trajectory dataset specifically designed for safety-critical scenarios. Comprehensive evaluations against several state-of-the-art models demonstrate that our model achieves average reduction of 8.8% in brierADE and 11.2% in brierFDE. A study based on human evaluation shows that predictions from the fine-tuned model better emphasizing rare behaviors. Overall, the proposed model maintains comparable performance on common behaviors while substantially improving the prediction of abrupt behavioral changes. This advancement enables autonomous vehicles to accurately anticipate pedestrian trajectories in safety-critical scenarios, thereby enhancing safety.