Research map / Multi-agent systems and games / Human-AI and goal-conditioned
LocoVR: Multiuser Indoor Locomotion Dataset in Virtual Reality
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.
Introduces LocoVR, a VR-collected dataset of 7000+ two-person indoor home trajectories across 131 scenes, and shows U-Net/Ynet models trained on it generalize better to unseen indoor scenes and real-world test data on three trajectory tasks.
Problem
Human trajectory datasets for indoor home environments are limited in scale and lack socially motivated multi-person navigation behaviors (proxemics), which hinders home robots and AI agents from predicting human locomotion.
Approach
Collect two-person locomotion data in VR by mapping tracked participants onto avatars in Habitat-Matterport scenes, enabling rapid scene variation and capture of social navigation behaviors, then train simple U-Net-based predictors on it for global path, trajectory, and goal prediction.
Key results
- A* + U-Net trained on LocoVR achieves mean Chamfer distance of 0.06/0.12/0.19 m across distance bins, outperforming A* + U-Net trained on GIMO (0.09/0.23/0.36) and THOR-MAGNI (0.07/0.21/0.30).
- U-Net trained on LocoVR reaches ADE of 0.11/0.24/0.44 m over 0-3 s horizons, versus 0.19/0.34/0.55 m for GIMO-trained U-Net.
- U-Net trained on LocoVR gets 72.2% object prediction accuracy at 0-3 m from goal versus 49.2% for GIMO-trained model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train U-Net on LocoVR rather than GIMO or THOR-MAGNI | U-Net trajectory prediction · LocoReal | ADE: 0.19, 0.34, 0.55 -> 0.11, 0.24, 0.44 | Helped |
| Train A* + U-Net on LocoVR rather than GIMO or THOR-MAGNI | A* + U-Net global path prediction · LocoReal | Mean Chamfer distance: 0.09, 0.23, 0.36 -> 0.06, 0.12, 0.19 | Helped |
| Train U-Net on LocoVR rather than GIMO or THOR-MAGNI | U-Net goal prediction · LocoReal | Object prediction accuracy: 49.2, 17.8, 3.9 -> 72.2, 40.1, 13.5 | Mixed |
| Reduce LocoVR to 19 scenes and 190 trajectories (data-size-G) (ablation) | LocoVR global path prediction · LocoReal | Mean Chamfer distance: 0.060, 0.119, 0.192 -> 0.077, 0.194, 0.326 | Hurt |
| Reduce LocoVR to 4 scenes and 658 trajectories (data-size-T) (ablation) | LocoVR global path prediction · LocoReal | Mean Chamfer distance: 0.060, 0.119, 0.192 -> 0.076, 0.163, 0.242 | Hurt |
| Remove the other person's movement input (wo/p2) (ablation) | LocoVR global path prediction · LocoReal | Mean Chamfer distance: 0.060, 0.119, 0.192 -> 0.061, 0.122, 0.205 | Hurt |
| Remove heading-direction input (wo/head) (ablation) | LocoVR global path prediction · LocoReal | Mean Chamfer distance: 0.060, 0.119, 0.192 -> 0.063, 0.134, 0.238 | Hurt |
| Reduce LocoVR to 19 scenes and 190 trajectories (data-size-G) (ablation) | LocoVR trajectory prediction · LocoReal | ADE: 0.111, 0.238, 0.441 -> 0.274, 0.496, 0.775 | Hurt |
| Reduce LocoVR to 4 scenes and 658 trajectories (data-size-T) (ablation) | LocoVR trajectory prediction · LocoReal | ADE: 0.111, 0.238, 0.441 -> 0.144, 0.297, 0.505 | Hurt |
| Remove the other person's movement input (wo/p2) (ablation) | LocoVR trajectory prediction · LocoReal | ADE: 0.111, 0.238, 0.441 -> 0.113, 0.254, 0.470 | Hurt |
| Remove heading-direction input (wo/head) (ablation) | LocoVR trajectory prediction · LocoReal | ADE: 0.111, 0.238, 0.441 -> 0.122, 0.254, 0.446 | Hurt |
| Reduce LocoVR to 19 scenes and 190 trajectories (data-size-G) (ablation) | LocoVR goal prediction · LocoReal | Goal position error: 0.83, 1.89, 3.45 -> 0.923, 2.165, 3.791 | Mixed |
Surprising findings
- Models trained on the much smaller GIMO dataset outperform THOR-MAGNI-trained models on the trajectory task, which the authors attribute to THOR-MAGNI's few scenes.
- Height and semantic scene maps outperformed binary obstacle maps for global path prediction, though the authors did not analyze why.
Limitations
- Virtual-to-real gap may affect walking speed and non-verbal behaviors such as facial expressions.
- Participants are able-bodied adults only, lacking age and motor-ability diversity.
- Avatar body motion from sparse trackers with FINAL-IK can display unnatural joint movements.
Experimental setup
Datasets: LocoVR GIMO THOR-MAGNI LocoReal HM3D
Benchmarks: LocoReal GIMO
Metrics: Chamfer distance (mean and max) ADE goal position error object prediction accuracy
Baselines: Ynet A* + MAP A* + DISTMAP RANDOM NEAREST
Builds on: Ynet A* Habitat-Matterport 3D
Closest papers by meaning
- RoomTour3D: Geometry-Aware Video-Instruction Tuning for Embodied Navigation (CVPR 2025 · similarity 0.87)
- Habitat 3.0: A Co-Habitat for Humans, Avatars, and Robots (ICLR 2024 · similarity 0.87)
- SocialCVAE: Predicting Pedestrian Trajectory via Interaction Conditioned Latents (AAAI 2024 · similarity 0.85)
- Social-Transmotion: Promptable Human Trajectory Prediction (ICLR 2024 · similarity 0.85)
- CityWalker: Learning Embodied Urban Navigation from Web-Scale Videos (CVPR 2025 · similarity 0.85)
- Image2Sim: Scaling Embodied Navigation via Generative Neural Simulator (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.