Mode-as-Sequence: New Framework for Multimodal Motion Prediction

A research paper on arXiv proposes Mode-as-Sequence, a unified decoding framework for multimodal motion forecasting. The framework translates an unordered mode set into an ordered mode sequence, explicitly modeling mode-to-mode dependency to address mode collapse and unreliable confidence ranking. Two instantiations are developed: ModeSeq, which performs recurrent mode decoding, and Parallel ModeSeq, which uses masked self-attention for parallel processing. The paper is available at arXiv:2605.24037.