Open-IO/humanoid-g1-odometry

Humanoid (Unitree G1) Inertial Odometry Dataset

Paper | Project Page | Code

SSIO (yellow), trained self-supervised on this dataset's 1 h of data, against SuperOdom ground truth (red), at 8× speed.

Real-robot recordings of a Unitree G1 humanoid: the pelvis IMU, all 29 joint states, and ground-truth pose and velocity from SuperOdom LiDAR-inertial odometry. It is the humanoid benchmark of SSIO, which trains inertial odometry from the robot's own unlabeled IMU stream.

Dataset Structure

Each clip is one folder, named <robot-id>_<category>_<YYYYMMDD>_<HHMMSS>, with the motions navigation, walking, combo, dance, kick, sprint, and general (humanoid).

Split	File	Clips	Frames	Duration (hours)
train	splits/train.txt	31	799,388	1.11
test	splits/test.txt	1	43,991	0.06
Total		32	843,379	1.17

The splits are the SSIO paper's. Validation uses the 31 training clips with non-overlapping windows, so it serves model selection only. The test clip, 10871_humanoid_20260520_223451, comes from the same session as 8 training clips.

Data Format

All arrays are plain NumPy files (allow_pickle=False), share the timestamps in time_ns.npy, and are finite.

• imus.npy (P, T, 9) float16: per body part, acc [0:3] (m/s², specific force, not bias-corrected), gyro [3:6] (rad/s), and ground-truth body-frame velocity gt_vel [6:9] (m/s). Cast to float32 before use.
• pose.npy (P, T, 7) float32: ground-truth world pose [tx, ty, tz, qx, qy, qz, qw]; the quaternion is body to world (R_WB), xyzw order.
• time_ns.npy (T,) int64: timestamps in ns at a nominal 200 Hz. The stream has gaps (up to 210 ms), so integrate with the real dt.
• joints.npy (T, 29, 3) float32: joint position q (rad), velocity dq (rad/s), and estimated torque tau_est (N·m), with names and units in joints.json.
• spec.json: format version, length T, and the part names.

P = 14 body parts, in the order of spec.json["names"]: pelvis, head, left_upper_arm, left_forearm, left_hand, right_upper_arm, right_forearm, right_hand, left_upper_leg, left_lower_leg, left_foot, right_upper_leg, right_lower_leg, right_foot. Part 0, pelvis, is the physical G1 IMU, re-expressed in the pelvis frame; the other 13 are IMU-format streams derived from the robot's body motion. Body frames report in the part's own frame; the world frame is z-up, with R_WB @ acc ≈ [0, 0, +g] at rest.

Quick Usage

pip install huggingface_hub numpy
hf download Open-IO/humanoid-g1-odometry --repo-type dataset --local-dir humanoid_g1

import json
from pathlib import Path
import numpy as np

clip = Path("humanoid_g1") / (Path("humanoid_g1/splits/train.txt").read_text().split()[0])
k = json.loads((clip / "spec.json").read_text())["names"].index("pelvis")

imus   = np.load(clip / "imus.npy")[k].astype(np.float32)   # (T, 9)
pose   = np.load(clip / "pose.npy")[k]                      # (T, 7)
t_ns   = np.load(clip / "time_ns.npy")                      # (T,)
joints = np.load(clip / "joints.npy")                       # (T, 29, 3)
acc, gyro, gt_vel_body = imus[:, 0:3], imus[:, 3:6], imus[:, 6:9]

To train or evaluate with the SSIO code, download into $DATA_ROOT/humanoid_g1 and use the HumanoidOdom_* dataset configs; the HumanoidOdom_Joints_* configs also read joints.npy.

Citation

https://ssio-imu.github.io

@misc{qiu2026ssioselfsupervisedinertialodometry,
      title={SSIO: Self-Supervised Inertial Odometry},
      author={Yuheng Qiu and Yutian Chen and Carlton Cort and Weison Huang and Shibo Zhao and Pieter Abbeel and Rocky Duan and Koushil Sreenath and C. Karen Liu and Carmelo Sferrazza and Junyi Geng and Sebastian Scherer and Guanya Shi},
      year={2026},
      eprint={2610.25988},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2610.25988},
}