Reproduce “TSDiT: Traffic Scene Diffusion Models With Transformers” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Chen Yang, Tianyu Shi

Cardiff University and University of Toronto researchers introduced TSDiT, a "world-centric" model integrating Diffusion models with Transformers to generate diverse and realistic traffic agent trajectories for autonomous driving simulations. The model achieved a lower Average Displacement Error (ADE) of 0.684 on the Waymo Motion Prediction dataset while enhancing the generation of smooth turning trajectories.

Key insights

  • A "world-centric" model paradigm effectively simplifies data preprocessing and enhances computational efficiency by enabling simultaneous multi-agent trajectory generation.
  • Integrating Diffusion models into trajectory generation significantly increases the diversity and stochasticity of agent actions, which is critical for robust simulation.
  • Dedicated Transformer modules (Other Agent Former, HD Map Former) are essential for capturing complex inter-agent interactions and ensuring adherence to map constraints like lanes and junctions.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.