Causally-Aware Hierarchical Graph Neural Networks for Robust Trajectory Prediction

Robust trajectory prediction of traffic participants is crucial for autonomous driving in complex urban environments. Traditional data-driven models predominantly rely on feature correlations and fail to explicitly model the causal relationships underlying traffic interactions. Motivated by this challenge, we propose a novel Causally-aware Hierarchical Graph Network, called CHGNet, for accurate and reliable vehicle trajectory prediction. The proposed framework enhances the robustness of autonomous vehicles by incorporating lane-aware scene context and jointly modeling both individual vehicle dynamics and inter-vehicle interactions. CHGNet introduces three key innovations: (i) a lane-aware spatio-temporal Transformer module that mitigates the impact of trajectory deviations and uncertainties from both temporal and spatial perspectives; (ii) a hierarchical graph neural network that effectively captures multi-level interaction patterns within the driving scene; and (iii) a causal reasoning module that explicitly models causal dependencies among vehicles while suppressing spurious correlations. Experimental results on the real-world dataset demonstrate that the proposed framework is resilient to imperfect tracklets and achieves competitive performance relative to state-of-the-art approaches, validating its effectiveness in improving predictive accuracy and robustness in complex traffic scenarios.