ARTFEED — Contemporary Art Intelligence

RankFormer: Transformer-Based Multi-Agent Trajectory Prediction

ai-technology · 2026-08-13

A new deep learning model, RankFormer, has been proposed for multi-modal trajectory prediction in autonomous driving and traffic safety. The model, detailed in arXiv paper 2604.07126, employs a pure Transformer architecture to handle temporal dependencies and agent-agent spatial interactions without relying on graph structures or manually labeled intentions. It uses two parallel decoders to generate trajectories and probabilities, addressing challenges such as complex decision-making and multiple possible intentions. The approach aims to improve prediction accuracy in dynamic traffic scenarios.

Key facts

  • RankFormer is a pure Transformer-based model for multi-modal trajectory prediction.
  • It addresses challenges in autonomous driving, traffic operations, and transportation safety analysis.
  • The model encodes historical trajectories and uses two parallel decoders for trajectory and probability generation.
  • It does not require graph structures or manually labeled intentions.
  • The paper is available on arXiv with identifier 2604.07126.
  • The model considers temporal dependencies and agent-agent spatial interactions.

Entities

Institutions

  • arXiv

Sources