Reproduce “TrafficBots: Towards World Models for Autonomous Driving Simulation and Motion Prediction” 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
Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, Luc Van Gool
TrafficBots introduces a differentiable traffic simulator formulated as a world model, generating realistic, configurable, and scalable multi-agent behaviors for autonomous driving planning. The system integrates a Transformer-based architecture, time-invariant latent personalities, and navigational destinations for bot agents, showing competitive performance on the Waymo Open Motion Dataset for joint interactive prediction and efficient real-time simulation.
Key insights
- Formulating data-driven traffic simulation as a differentiable "world model" enables the creation of more sophisticated and trainable simulation environments for autonomous driving planning.
- Explicitly conditioning agent behaviors on time-invariant latent "personality" and long-term "navigational destination" is crucial for generating diverse, configurable, and causally sound multi-modal behaviors in closed-loop simulation.
- Efficient contextual representation, including a novel Angular Encoding for directional information, and a shared policy among agents are vital for scalable and faster-than-real-time simulation in dense, multi-agent urban environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.