Reproduce “SIMPL: A Simple and Efficient Multi-agent Motion Prediction Baseline for Autonomous Driving” on OpenResearch

About this paper

Lu Zhang, Peiliang Li, Sikang Liu, Shaojie Shen

The SIMPL (Simple and Efficient Multi-agent Motion Prediction) framework, developed by HKUST and DJI, addresses the efficiency-accuracy trade-off in motion prediction by introducing an instance-centric scene representation and a symmetric fusion transformer, achieving competitive prediction performance with real-time efficiency. The method further represents trajectories as smooth Bézier curves, providing kinematically consistent outputs suitable for autonomous driving planning modules.

Key insights

• An instance-centric scene representation, computing relative spatial relationships between all semantic elements, maintains spatial fidelity while avoiding the computational redundancy of traditional agent-centric methods.
• A Symmetric Fusion Transformer (SFT) uses directed multi-head cross-attention and updated relative positional embeddings to achieve viewpoint-invariant feature fusion for simultaneous multi-agent prediction.
• Representing future trajectories as Bézier curves provides inherent smoothness and direct access to higher-order derivatives, which is crucial for downstream planning modules and ensures numerical stability.

Summary by alphaXiv.