<ul><li>Traffic simulation aims to learn a policy for traffic agents that, when unrolled in closed-loop, faithfully recovers the joint distribution of trajectories observed in the real world.</li><li>Tokenized multi-agent policies have become the state-of-the-art in traffic simulation, but they suffer from covariate shift when executed in closed-loop during simulation.</li><li>A new strategy called Closest Among Top-K (CAT-K) rollouts is presented to mitigate covariate shift, enabling improved performance of tokenized traffic simulation policies.</li><li>CAT-K fine-tuning outperforms larger models in the Waymo Sim Agent Challenge leaderboard, achieving the top spot.</li></ul>

Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models

Discover more