EgoTraj dataset brings 75 real-world egocentric trajectories for AI prediction
New open dataset from Meta Quest Pro captures gaze and head motion for humanoid robotics
A team of researchers from multiple institutions has released EgoTraj, an open-source egocentric human trajectory dataset designed to advance multimodal prediction for applications like humanoid robotics and wearable AR. Recorded using Meta Quest Pro headsets, the dataset includes 75 long-horizon, self-directed navigation sequences from diverse participants in real-world urban environments. Each sequence provides synchronized RGB video, continuous 6-degree-of-freedom head poses, per-frame 3D eye gaze vectors, and scene annotations—a combination rarely available in existing datasets.
Benchmarking several state-of-the-art trajectory prediction methods, the team conducted ablation studies to isolate contributions from gaze, scene, and motion cues. Results highlight significant improvements in prediction accuracy when integrating egocentric gaze and scene context. The EgoTraj dataset, along with tools like the EgoViz Dashboard, is publicly available to accelerate research in AR perception, navigation aids, and assistive systems.
- 75 real-world urban navigation sequences recorded from multiple Meta Quest Pro wearers
- Includes synchronized RGB video, 6-DoF head poses, 3D eye gaze, and scene annotations
- Benchmarked state-of-the-art methods with ablation studies showing gaze and scene cues improve trajectory prediction
Why It Matters
Egocentric trajectory data is critical for humanoid robots and AR glasses to anticipate and assist human movement in real time.