You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset contains egocentric recordings of public and semi-public indoor spaces. Faces were automatically detected and blurred with EgoBlur Gen2, but automated anonymization is not perfect. By requesting access you agree to use the data for non-commercial research only, to make no attempt to identify any individual appearing in it, and not to redistribute it.
Log in or Sign Up to review the conditions and access this dataset content.
EgoNav Dataset
Egocentric navigation data collected with a chest-mounted sensor rig (the "Smartbelt") walking through indoor and campus environments. Each episode pairs a continuous 6-DoF trajectory with synchronized panoramic and camera imagery, plus DINOv3 features precomputed over the camera stream.
This is the exact data used to train the diffusion trajectory model in EgoNav: Learning Humanoid Navigation from Human Data (project page, IEEE RA-L 2026).
Contents
| File | Shape | dtype | What it is |
|---|---|---|---|
<episode>_da |
(N, 25) |
float32 | Trajectory: time, position, orientation, velocity, indices |
<episode>_pano_uint8 |
(M, 5, 140, 360) |
uint8 | Panorama: RGB + depth + semantic class |
<episode>_dinov3_patch_fp16 |
(K, 15, 27, 384) |
float16 | DINOv3 ViT-S/16 patch features over the camera frames |
<episode>_rgb.zip |
K × (480, 848, 3) |
JPEG q85 | Anonymized camera frames |
<episode>_depth.npz |
(J, 480, 848, 1) |
float32 | Depth — sparse, see below |
<episode>_video_t.npy / _pano_t.npy |
float64 | Sensor timestamps (seconds) | |
<episode>.done.json |
Per-episode build metadata |
_da, _pano_uint8 and _dinov3_patch_fp16 are raw memory-mapped arrays —
open them with np.memmap / np.load(..., mmap_mode='r') using the shapes in
.done.json, or just use the reader below.
Trajectory columns (_da)
0 time (s, relative to episode start)
1:4 position (x, y, z)
4:8 orientation quaternion
8:11 pose variance
11:14 linear velocity
14:17 angular velocity
17 step
18:22 joint values
22 point-cloud index
23 panorama index -> indexes _pano_uint8
24 camera index -> indexes _rgb.zip / _dinov3_patch_fp16
Panorama channels (_pano_uint8)
[R, G, B, depth, semantic]. RGB and depth are 0–255 (255 depth = unknown).
Semantic is an integer class:
0 ground/sidewalk 1 stairs 2 door 3 wall/pillar
4 furniture/objects 5 person 6 grass/dirt 7 unlabeled
Downloading
The two levels are separate directories, so you can take only what you need.
from huggingface_hub import snapshot_download
# training data only (~163 GB)
snapshot_download("ASKKER/egonav", repo_type="dataset", local_dir="egonav",
allow_patterns=["level1/*", "*.json", "*.py"])
# add the raw sensor layer (~26 GB) when you want to run your own encoder
snapshot_download("ASKKER/egonav", repo_type="dataset", local_dir="egonav",
allow_patterns=["level2/*"])
# or a single episode
snapshot_download("ASKKER/egonav", repo_type="dataset", local_dir="egonav",
allow_patterns=["*eDS20HZVZS_V4Data_260205gatesbase*"])
Reading it
from egonav_data import Episode, list_episodes
print(list_episodes("dataset"))
ep = Episode("dataset", "eDS20HZVZS_V4Data_260205gatesbase")
ep.traj # (N,25) float32
ep.col("pos") # (N,3) position
ep.pano[i] # (5,140,360) uint8
ep.dino[i] # (15,27,384) float16
ep.rgb(i) # (480,848,3) uint8, decoded from JPEG
ep.rgb_for_step(t) # camera frame aligned to trajectory row t
Only numpy is required; rgb() decodes JPEG with OpenCV if available (falls back to Pillow). Nothing else.
Sensor sampling is close to, but not perfectly, uniform — prefer the raw
video_t / pano_t timestamps over assuming a fixed FPS.
Citation
@article{wang2026egonav,
title={Learning Humanoid Navigation from Human Data},
author={Wang, Weizhuo and Ze, Yanjie and Liu, C. Karen and Kennedy III, Monroe},
journal={IEEE Robotics and Automation Letters},
year={2026},
}
- Downloads last month
- 115