Research map / Label-efficient and robust learning / Self-supervised and contrastive

UniTraj: Learning a Universal Trajectory Foundation Model from Billion-Scale Worldwide Traces

NeurIPS 2025 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Self-supervision in supervised pipelines · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Builds WorldTrace, a 2.45M-trajectory global GPS dataset from OSM traces, and UniTraj, an encoder-decoder trajectory foundation model pre-trained with adaptive resampling and masked trajectory reconstruction, evaluated on recovery, prediction and classification.

Problem

Existing trajectory models are task-specific, region-dependent and sensitive to data quality, and no large, globally diverse trajectory dataset or universal trajectory foundation model exists.

Approach

Pre-train a masked autoencoder on lat/lon/time only (no POI or road network) over a worldwide trajectory corpus, using logarithmic length-dependent resampling and four masking strategies (random, block, key-points, last-N) so one backbone serves multiple tasks zero-shot or after fine-tuning.

Key results

• Trajectory recovery: fine-tuned UniTraj reaches MAE 6.94 m on WorldTrace versus 47.64 for the best baseline TrajFM, and 23.23 vs 34.53 for TrajBERT on GeoLife.
• Trajectory prediction: zero-shot UniTraj MAE 49.85 on WorldTrace, a 10.49% improvement over the best baseline; fine-tuning gives 30.10.
• Classification: zero-shot 71.3% accuracy on GeoLife and 79.3% after fine-tuning on Grab-Posisi.

Measured effects reported in the paper

Surprising findings

• Removing block masking gives the best MAE on WorldTrace, Chengdu, Xi'an and Grab-Posisi, and removing key-points masking improves Porto, so individual strategies do not uniformly help.
• Increasing WorldTrace from 1M to 2.45M trajectories slightly increased MAE, attributed to added noise, while a 1M high-quality subset performed better.

Limitations

• Excludes region-bound context (POI, road network) by design, so baselines relying on such information were not compared.
• Dataset noise: the full 2.45M set degrades performance relative to a curated subset.
• Trajectories are restricted to motorized movement from OSM traces.

Experimental setup

Datasets: WorldTrace Chengdu Xi'an GeoLife Grab-Posisi Porto

Benchmarks: WorldTrace Chengdu Xi'an GeoLife Grab-Posisi Porto

Metrics: MAE RMSE accuracy

Baselines: Linear DHTR Transformer DeepMove TrajBERT TrajFM

Builds on: masked autoencoder rotary position encoding Ramer-Douglas-Peucker algorithm

Closest papers by meaning

• STraj: Self-training for Bridging the Cross-Geography Gap in Trajectory Prediction (AAAI 2025 · similarity 0.86)
• Trajectory World Models for Heterogeneous Environments (ICML 2025 · similarity 0.86)
• TransferTraj: A Vehicle Trajectory Learning Model for Region and Task Transferability (NeurIPS 2025 · similarity 0.85)
• GTR: A General, Multi-View, and Dynamic Framework for Trajectory Representation Learning (ICML 2025 · similarity 0.85)
• TrajAgent: An LLM-Agent Framework for Trajectory Modeling via Large-and-Small Model Collaboration (NeurIPS 2025 · similarity 0.84)
• SingularTrajectory: Universal Trajectory Predictor Using Diffusion Model (CVPR 2024 · similarity 0.84)