Hierarchical Transformers for Motion Forecasting Based on Inverse Reinforcement Learning

Understanding and forecasting the future states of surrounding agents is crucial for ensuring safe and human-like decision-making for autonomous vehicles, which plays a pivotal role in the future of intelligent transportation systems. While data-driven methodologies, including recent advancements such as transformers and graph neural networks, have made significant progress, managing multimodality, achieving scene conformability, and providing human-like decision-making in future steps of the autonomy stack remains challenging. Unlike existing data-driven motion forecasting approaches that do not explicitly consider human-like motion planning, this work proposes learning from human (expert) behavior based on Inverse Reinforcement Learning (IRL). We design a novel Hierarchical Transformer-based Motion Forecasting framework, called HTMF, that learns the multimodal future motion distribution using maximum entropy IRL and leverages transformer-based learned occupancy grid maps as a context-aware approximation to the true distribution, enabling the generation of realistic motions. To assess the performance of the proposed approach, we conduct comprehensive experiments on large-scale real-world datasets across diverse challenging scenarios such as highways and unsignalized intersections. The results demonstrate that HTMF outperforms other benchmark motion forecasting methods in terms of prediction errors and in generating more plausible and human-like trajectories.