Enhancing Consistency Models for Multi-Agent Trajectory Prediction

TL;DR

This paper introduces ECTraj, an enhanced consistency model framework for multi-agent trajectory prediction that achieves faster inference and higher accuracy, setting new benchmarks on Argoverse 2.

Contribution

The paper develops an improved training and conditional generation method for consistency models, enabling efficient one-step trajectory prediction in autonomous driving.

Findings

ECTraj achieves faster inference with competitive accuracy.

The framework establishes new benchmarks on the Argoverse 2 dataset.

Enhanced training improves the quality of multi-agent trajectory predictions.

Abstract

Diffusion models for multi-agent trajectory prediction are limited by iterative denoising, which causes inference latency that hinders their use in time-critical settings like autonomous driving. Fast-sampling variants using DDIM and informed initial noise distributions partially alleviate this issue, but they either fail to achieve true single-step generation or are constrained by the chosen noise distribution. Consistency Models (CMs) offer high-quality one-step generation by mapping noise directly to data, but are difficult to train from scratch . We propose ECTraj, an enhanced CM pipeline with improved training and conditional generation for trajectory prediction. Our framework extends the student-teacher consistency training scheme: the student produces standard outputs, while the teacher explicitly fuses its predictions with parts of the ground truth to give stronger supervision. We also exploit CMs' direct denoising for top-K multi-shot generation during training. Combining conditional generation with this enhanced consistency objective yields faster inference and improved prediction accuracy, establishing competitive new benchmarks on the large-scale Argoverse 2 dataset.