MHD-Planner: High-Efficiency Motion Planning for Autonomous Driving with Hybrid Mamba-Attention

Generative diffusion models are powerful tools for motion planning, yet their reliance on Transformer backbones with quadratic complexity hinders real-time application. To overcome this bottleneck, we propose MHD-Planner, centered on a novel hybrid Mamba-Attention architecture designed for both efficiency and high-performance. Realized in our PrismNet denoising network, this architecture employs a bidirectional Mamba (Hydra SSM) for linear-time temporal modeling, reserving cross-attention for fusing rich scene context. This synergistic design proves its effectiveness on the nuPlan benchmark, where MHD-Planner not only achieves a 2× inference speedup over a Transformer baseline but also delivers superior planning performance in demanding reactive simulations. Consequently, our work demonstrates that the proposed hybrid architecture offers a compelling solution to the efficiency-performance tradeoff, paving the way for real-time, high-fidelity motion planning in autonomous driving.