wrl2003/Humanoid-OmniOCC

Humanoid OmniOcc RGB, State, and Occupancy Dataset

This dataset contains simulated and real-world Humanoid OmniOcc data. It primarily provides:

• Simulated RGB images;
• Robot state, action, and camera-pose metadata;
• Semantic occupancy grids;
• Real-world stereo images, merged point clouds, and occupancy grids.

The simulated subset contains 1,367 episodes, and the real-world subset contains 25 scenes. To support selective downloading and reliable transfer of this large dataset, every simulated episode and real-world scene is stored in a separate .tar.zst archive.

Repository Layout

Humaniod_omniocc_rgb_common_occ/
|-- sim/
|   |-- <scene_name>/
|   |   |-- <episode_timestamp>.tar.zst
|   |   `-- ...
|   `-- ...
|-- real/
|   |-- 4f_ba1.tar.zst
|   |-- 4f_ba2.tar.zst
|   |-- ...
|   `-- jiudian8.tar.zst
`-- _meta/
    |-- SHA256SUMS.txt
    |-- uploaded_archives.tsv
    |-- uploaded_success.txt
    |-- clean_sim_upload_map.tsv
    |-- README_RESTORE.txt
    `-- real/
        |-- SHA256SUMS.txt
        |-- uploaded_archives.tsv
        |-- uploaded_success.txt
        |-- humanoid_upload_path_map.txt
        `-- README_RESTORE.txt

The _meta/ directory contains checksums, source-to-repository path mappings, and upload completion records. All archives under sim/ and real/ have been uploaded successfully.

Simulated Data

Each sim/<scene_name>/<episode_timestamp>.tar.zst archive contains one complete episode. The scene name and episode timestamp are preserved after extraction:

<scene_name>/
`-- <episode_timestamp>/
    `-- state/
        |-- rgb/
        |   `-- rgb/
        |       |-- robot.front_camera.left.rgb_image/
        |       |   |-- 00000030.jpg
        |       |   `-- ...
        |       |-- robot.front_camera.right.rgb_image/
        |       |-- robot.zed2i_front.left.rgb_image/
        |       |-- robot.zed2i_front.right.rgb_image/
        |       |-- robot.zed2i_left.left.rgb_image/
        |       |-- robot.zed2i_left.right.rgb_image/
        |       |-- robot.zed2i_right.left.rgb_image/
        |       |-- robot.zed2i_right.right.rgb_image/
        |       |-- robot.zed2i_back.left.rgb_image/
        |       `-- robot.zed2i_back.right.rgb_image/
        |-- common/
        |   |-- 00000030.npy
        |   `-- ...
        `-- occ_gt_lable/
            |-- 00000030_occupancy_gt.npy
            `-- ...

Note that state/rgb/rgb/ is the actual directory hierarchy in the source data. The directory name occ_gt_lable is also preserved as originally provided.

RGB Images

• Format: JPEG;
• Resolution: 960 x 600;
• The numeric file stem is the frame ID;
• Ten camera streams are provided: a front stereo camera and four ZED2i stereo pairs facing front, left, right, and back.

Common State and Metadata

Each state/common/<frame_id>.npy is a NumPy object scalar. Load it with allow_pickle=True:

import numpy as np

common = np.load("00000030.npy", allow_pickle=True).item()

The main fields are:

Field	Shape / dtype	Description
target_path	(N, 2), float64	Target path
robot.action	(2,), float64	Robot action
robot.position	(3,), float32	Robot position
robot.orientation	(4,), float32	Robot orientation quaternion
robot.joint_positions	(19,), float32	Joint positions
robot.joint_velocities	(19,), float32	Joint velocities
<camera>.position	(3,), float32	Camera position
<camera>.orientation	(4,), float32	Camera orientation quaternion
<camera>.segmentation_info	dict	Semantic segmentation metadata
<camera>.instance_id_segmentation_info	dict	Instance segmentation metadata

The length N of target_path may vary between frames. Always use the array's actual shape.

Occupancy Grids

Each state/occ_gt_lable/<frame_id>_occupancy_gt.npy has:

• Shape: (384, 384, 44);
• dtype: uint8;
• Values representing occupancy or semantic label IDs rather than RGB intensities;
• Possible value 255, which should be handled as a special or ignored label according to the downstream task.

Semantic visualization PNG files and large PLY files from the simulated source data are intentionally excluded. This repository retains the *_occupancy_gt.npy files needed for training.

Real-World Data

Each real/<scene_name>.tar.zst archive extracts to one scene directory:

<scene_name>/
|-- 000000_cam0_left.jpg
|-- 000000_cam0_right.jpg
|-- 000000_cam1_left.jpg
|-- 000000_cam1_right.jpg
|-- 000000_cam2_left.jpg
|-- 000000_cam2_right.jpg
|-- 000000_cam3_left.jpg
|-- 000000_cam3_right.jpg
|-- <scene_name>_merged.npy
|-- <scene_name>_merged.pcd
`-- occ_output/
    |-- occupancy_gt.npy
    `-- occupancy_gt.ply

Real-world file formats:

File	Format
*_cam*_left.jpg, *_cam*_right.jpg	JPEG images from four stereo camera pairs
<scene_name>_merged.npy	Merged point cloud with shape (N, 4) and columns x, y, z, intensity
<scene_name>_merged.pcd	Binary PCD v0.7 with fields x, y, z, intensity
occ_output/occupancy_gt.npy	uint8 occupancy grid with shape (384, 384, 44)
occ_output/occupancy_gt.ply	ASCII PLY visualization with fields x, y, z, red, green, blue

The point count N varies by scene.

Real-World Annotation Pipeline and Quality Assurance

The real-world occupancy ground truth is not generated by directly voxelizing raw LiDAR scans. Instead, the annotation pipeline consists of three stages:

1. Accurate LiDAR-camera calibration followed by multi-frame LiDAR accumulation;
2. Visibility-aware occupancy generation through geometric reasoning;
3. Semi-automatic annotation followed by manual verification and correction of potential annotation errors.

LiDAR-Camera Calibration Quality

Accurate LiDAR-camera calibration is essential because any extrinsic error would introduce systematic offsets between the accumulated geometry and the camera observations, and those offsets would propagate into the generated occupancy annotations. In the examples below, the projected LiDAR points closely follow the corresponding RGB surfaces and object boundaries across multiple viewpoints. The consistent overlap on walls, desks, chairs, monitors, and other foreground objects demonstrates the geometric consistency of the LiDAR measurements with the camera views and indicates that the calibration is accurate.

Representative LiDAR-to-camera projection results. The close alignment between projected points and RGB structures provides a direct qualitative check of the calibration quality.

Occupancy Annotation Visualizations

After calibration and multi-frame accumulation, visibility-aware geometric reasoning is used to generate occupancy annotations. The results are then refined through semi-automatic annotation and manually verified to correct potential errors. The following examples visualize the resulting semantic occupancy annotations projected into camera views. Their agreement with room layout and object extents provides an additional qualitative check of the annotation pipeline.

Optional Depth Data

The depth images are very large: approximately 6.5 TB in decimal units, or about 6.0 TiB. Depth is not required to use the RGB, common state, and occupancy data in this repository. Download it only when per-frame depth is needed for your task.

The complete depth data is available as a separate dataset:

Humaniod OmniOcc Depth

Its <scene_name>/<episode_timestamp> identifiers match the simulated episodes in this repository:

sim_depth/
`-- <scene_name>/
    `-- <episode_timestamp>.tar.zst

The main extracted path is:

<scene_name>/<episode_timestamp>/state/depth/depth/<camera_name>/<frame_id>.npy

Each depth frame has:

• Shape: (600, 960);
• dtype: float32;
• The same frame ID and camera naming convention as its corresponding RGB image.

Quick Samples

The samples/ directory provides one complete frame from both subsets without requiring users to download a full archive:

• samples/sim/: frame 00000030 from scene AIUE_V01_001, episode 2025-10-18_14-49-48-685163;
• samples/real/: frame 000000 and all associated point-cloud and occupancy files from scene 4f_ba1.

Simulated front-camera RGB:

Real-world camera RGB:

See samples/README.md for the full sample layout and NumPy loading examples.

Download and Extraction

After installing and logging in with the ModelScope CLI, download an individual archive with:

ms download \
  ruilin.wang/Humaniod_omniocc_rgb_common_occ \
  sim/AIUE_V01_001/2025-10-18_14-49-48-685163.tar.zst \
  --repo-type dataset \
  --local-dir ./humanoid_omniocc

Extract a .tar.zst archive with:

tar --use-compress-program=unzstd \
  -xf 2025-10-18_14-49-48-685163.tar.zst

Alternatively:

zstd -dc 2025-10-18_14-49-48-685163.tar.zst | tar -xf -

Verify downloaded simulated archives against the SHA-256 manifest:

sha256sum -c _meta/SHA256SUMS.txt

For real-world archives, use _meta/real/SHA256SUMS.txt.

Matching Data Across Modalities

Simulated RGB, common state, occupancy, and the separate depth dataset are matched by:

<scene_name>/<episode_timestamp>

Within an episode, modalities are aligned by the numeric frame ID:

common/00000030.npy
rgb/rgb/<camera_name>/00000030.jpg
occ_gt_lable/00000030_occupancy_gt.npy
depth/depth/<camera_name>/00000030.npy