Map-Free Trajectory Prediction via Dual-Path Spatial-Temporal Network with Mamba

Trajectory prediction is essential for autonomous driving, enabling vehicles to anticipate the future motions of surrounding agents. However, most existing approaches rely on high-definition (HD) maps, which are expensive to obtain and often unavailable in fast-changing regions. To overcome this limitation, we propose DSTM (Dual-path Spatial-Temporal Network with Mamba), a map-free prediction framework that separately models agent dynamics and inter-agent relational evolution. DSTM employs a dual-path encoder: the Temporal Motion Branch leverages Mamba's efficient long-range sequence modeling and attention mechanisms to capture individual motion behaviors and spatial context, while the Spatial-Temporal Relation Branch encodes geometric relations and learns their evolution patterns using Temporal Evolution Mamba. Subsequently, the agent-to-agent interaction is inferred with graph neural network, and the fused features are decoded into multimodal trajectories. Experiments on the Argoverse and INTERACTION datasets demonstrate that DSTM outperforms existing map-free baselines, reducing minADE, minFDE, and miss rate by 16.3%, 19.7%, and 24.4%, respectively, compared to CRAT-Pred on Argoverse, and remains competitive with map-based models. Compared to a Transformer-based variant under the same dual-path architecture, DSTM reduces computational cost by 36% MACs and 37% parameters without sacrificing prediction accuracy.