MetaTra: Meta-Learning for Generalized Trajectory Prediction in Unseen Domain

TL;DR

MetaTra tackles the challenge of domain generalization in trajectory prediction by integrating a Dual Trajectory Transformer that captures both individual intent and group interactions with a meta-learning framework (MGTP) designed to simulate domain shifts between source and unseen target domains. The MGTP framework is augmented with Serial and Parallel Training (SPT) for stability and MetaMix for feature augmentation, all within a CVAE-based probabilistic predictor to generate multimodal futures. Across ETH-UCY, SDD, and NBA datasets, MetaTra achieves state-of-the-art generalization and can serve as a plug-in to improve existing predictors, with notable cross-domain transfer gains such as NBA→Soccer. The approach demonstrates that combining tailored architectural design for spatio-temporal interactions with meta-learning and data augmentation can robustly extend trajectory prediction to unseen environments, which is crucial for autonomous navigation and robotic planning.

Abstract

Trajectory prediction has garnered widespread attention in different fields, such as autonomous driving and robotic navigation. However, due to the significant variations in trajectory patterns across different scenarios, models trained in known environments often falter in unseen ones. To learn a generalized model that can directly handle unseen domains without requiring any model updating, we propose a novel meta-learning-based trajectory prediction method called MetaTra. This approach incorporates a Dual Trajectory Transformer (Dual-TT), which enables a thorough exploration of the individual intention and the interactions within group motion patterns in diverse scenarios. Building on this, we propose a meta-learning framework to simulate the generalization process between source and target domains. Furthermore, to enhance the stability of our prediction outcomes, we propose a Serial and Parallel Training (SPT) strategy along with a feature augmentation method named MetaMix. Experimental results on several real-world datasets confirm that MetaTra not only surpasses other state-of-the-art methods but also exhibits plug-and-play capabilities, particularly in the realm of domain generalization.

Figures (8)

• Figure 1: In part (a), we quantify the statistical biases in various scenarios within the SDD. The terms "Density” ,"Speed”, and "Acceleration” refer to the average pedestrian count, velocity, and acceleration metrics for each sequence, respectively. In part (b), our approach to the generalized trajectory prediction challenge is presented. It shows three source domains on the left for training, each with distinct trajectory distributions, and some unfamiliar target domains on the right for testing. Our model acquires domain-invariant universal knowledge by simulating domain shifts through meta-train and meta-test, thereby effectively generalizing to unseen domains.
• Figure 2: Model architecture of the proposed Dual Trajectory Transformer (Dual-TT).
• Figure 3: Meta-learning framework for Generalized Trajectory Prediction (MGTP).
• Figure 4: Detail analysis for ETH-UCY and SDD.
• Figure 5: Performance metrics for 5 NBA teams.