video video 3 6 | label class label 312
classes |
|---|---|
01 | |
01 | |
110 | |
110 | |
1211 | |
1211 | |
2312 | |
2312 | |
3413 | |
3413 | |
4514 | |
4514 | |
5615 | |
5615 | |
6716 | |
6716 | |
7817 | |
7817 | |
8918 | |
8918 | |
10019 | |
10019 | |
1112 | |
1112 | |
11220 | |
11220 | |
12321 | |
12321 | |
13422 | |
13422 | |
14523 | |
14523 | |
15624 | |
15624 | |
16725 | |
16725 | |
17826 | |
17826 | |
18927 | |
18927 | |
20028 | |
20028 | |
21129 | |
21129 | |
2223 | |
2223 | |
22330 | |
22330 | |
2464 | |
2464 | |
2575 | |
2575 | |
2686 | |
2686 | |
2797 | |
2797 | |
2908 | |
2908 | |
3019 | |
3019 | |
01 | |
01 | |
110 | |
110 | |
2100 | |
2100 | |
3101 | |
3101 | |
4102 | |
4102 | |
5103 | |
5103 | |
6104 | |
6104 | |
7105 | |
7105 | |
8106 | |
8106 | |
9107 | |
9107 | |
10108 | |
10108 | |
11109 | |
11109 | |
1211 | |
1211 | |
13110 | |
13110 | |
14111 | |
14111 | |
15112 | |
15112 | |
16113 | |
16113 | |
17114 | |
17114 | |
18115 | |
18115 | |
19116 | |
19116 |
SixthSense G1 Contact Dataset
Real-world proprioceptive recordings and paired camera videos for studying whole-body contact sensing across standing, walking, and lying scenarios.
It contains Unitree G1 robot signals, short recording windows, dual-camera videos, and scenario-specific contact annotations. The dataset accompanies SixthSense: Proprioception-Only Whole-Body Wrench Estimation for Humanoids (arXiv:2605.01427, accepted by IEEE Robotics and Automation Letters, RA-L).
Dataset Overview
| Scenario | Recordings | Frames per recording | Duration | Force reference |
|---|---|---|---|---|
| Standing | 312 | 150 | 3 s | Sensor-based force reference and annotated contact link |
| Walking | 30 | 300 | 6 s | Not available |
| Lying | 30 | 300 | 6 s | Not available |
| Total | 372 | 64,800 frames | 21.6 min of robot signals |
Robot signals are sampled at 50 Hz. The complete release is approximately 0.62 GB. Video timing information is stored in each recording; camera frame rates and timing estimates are separate from the robot sampling rate. Timestamp banners are removed from both camera views using a fixed top crop of approximately 2.4% (16 pixels). The image is uniformly scaled and the horizontal edges are center-cropped to preserve each video's original resolution and aspect ratio. Frame counts, frame rates, and temporal alignment are preserved; original timing metadata remains in the HDF5 files.
UI Preview
The demo shows synchronized camera videos, contact force, IMU signals, and joint angles, velocities, and motor torques.
The preview shows standing recording g01_pair_001 (torso contact): both cameras, the contact-force reference, IMU signals, and all 29 joint angles, velocities, and motor torques. The GIF loops the 3-second recording at 0.5× speed. This synchronized visualization is rendered directly from the original recording and timestamps.
File Structure
README.md
LICENSE
dataset-demo.gif
stand/
manifest.json
annotations.json
1/
data.h5
camera_1.mp4
camera_2.mp4
...
walk/
manifest.json
1/ ... 30/
lie/
manifest.json
1/ ... 30/
Each numbered directory is one recording with an HDF5 signal file and two camera videos. Scenario manifests contain recording metadata and SHA-256 checksums.
Use scenario together with pair_id to identify a recording. Numeric directory names are browsing indices and may change between releases. Standing recordings are grouped by contact link, with torso followed by pelvis.
Signal Fields
T denotes the number of frames in a recording.
| HDF5 field | Shape | Description |
|---|---|---|
time_s |
[T] |
Recording time in seconds |
time_from_event_s |
[T] |
Standing-only time relative to the annotated event |
robot/joint_names |
[29] |
Joint names in signal-column order |
robot/q |
[T, 29] |
Joint angle, rad |
robot/qdot |
[T, 29] |
Joint velocity, rad/s |
robot/tau |
[T, 29] |
Motor output torque signal, N·m |
robot/imu |
[T, 10] |
Quaternion [x, y, z, w], angular velocity [x, y, z] in rad/s, acceleration [x, y, z] in m/s² |
contact/contact_link_name |
scalar | Annotated contact link, or unknown |
contact/force_valid |
[T] |
Validity mask for the force reference |
contact/force_pelvis_N |
[T, 3] |
Standing-only force reference in pelvis axes, N |
contact/moment_about_link_origin_pelvis_Nm |
[T, 3] |
Standing-only link-origin moment reference in pelvis axes, N·m |
replay/qpos, replay/qvel |
[T, 36], [T, 35] |
MuJoCo replay state |
Robot replay fields are optional for data inspection. Robot geometry and scene descriptions are supplied by the companion code repository.
Additional fields contain native-rate signals, source timestamps, validity masks, standing event metadata, and camera timing information. Motor torque is a robot signal, not an external contact-torque measurement.
Read a Recording
python -m pip install h5py
import h5py
with h5py.File("stand/1/data.h5", "r") as f:
time = f["time_s"][:]
joint_angle = f["robot/q"][:]
joint_velocity = f["robot/qdot"][:]
motor_torque = f["robot/tau"][:]
imu = f["robot/imu"][:]
metadata = f["metadata_json"].asstr()[()]
if "contact/force_pelvis_N" in f:
force = f["contact/force_pelvis_N"][:]
force_valid = f["contact/force_valid"][:]
Annotation and Evaluation Notes
Standing force vectors are constructed from a baseline-corrected single-axis force sensor and annotated contact direction. They are not independently measured three-axis forces. Link-origin moments are geometric references derived from force and contact position, rather than independent six-axis torque measurements.
Walking and lying provide robot signals and paired videos, but no measured force reference or contact onset/end labels. Use time_s for time relative to the recording start; their camera frame tables use the same convention.
Existing train/validation/test assignments are preserved where provided; standing recordings have no split assignment in the current release. When defining evaluation splits, group recordings by recording_family or source session to avoid splitting correlated recordings across train and test.
Replay is kinematic. Unmeasured base translation is held fixed, and camera synchronization may use estimated timing. Some observation conventions are inferred from the recording pipeline. Consult the per-recording metadata and validity masks before quantitative evaluation.
Citation
Please cite the arXiv preprint below. The paper has been accepted by RA-L. The entry will be updated with the journal DOI and bibliographic details when available.
@article{chen2026sixthsense,
title = {SixthSense: Task-Agnostic Proprioception-Only Whole-Body Wrench Estimation for Humanoids},
author = {Chen, Xingzhou and Xu, Xiayan and Ning, Yan and Shi, Ling and Yu, Jiyu and Zhang, Yizheng and Qian, Siyi and Xiang, Lingzhu and Chen, Jiahao and Wang, Yuquan and Zhang, Haodong},
journal = {arXiv preprint arXiv:2605.01427},
year = {2026},
doi = {10.48550/arXiv.2605.01427},
url = {https://arxiv.org/abs/2605.01427},
note = {Accepted for publication in IEEE Robotics and Automation Letters (RA-L)}
}
License
This dataset is licensed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). Non-commercial use, adaptation, and redistribution are permitted with appropriate attribution. Commercial use requires separate permission from the authors. Third-party robot assets remain subject to their upstream licenses.
- Downloads last month
- 369
