The article presents MotionLM, a model for multi-agent motion prediction used in autonomous vehicles. Unlike previous models, MotionLM represents continuous trajectories as sequences of discrete motion tokens, treating the prediction task as a language modeling problem. This approach eliminates the need for anchors or explicit latent variable optimization. Instead, the model maximizes the average log probability over sequence tokens. The model also avoids post-hoc interaction heuristics, producing joint distributions over interactive agent futures in a single autoregressive decoding process. The model’s sequential factorization allows for temporally causal conditional rollouts. The study shows that this approach has set new performance standards on the Waymo Open Motion Dataset.
Publication date: 28 Sep 2023
Project Page: https://arxiv.org/abs/2309.16534v1
Paper: https://arxiv.org/pdf/2309.16534
