Moving: 1,000 G1 locomotion motions

This dataset contains 1,000 accepted, nominally 10-second Unitree G1 motion clips generated with a pretrained SONIC planner and whole-body tracking controller in headless MuJoCo. Each clip contains both the measured simulated motion and the kinematic reference, recorded at approximately 50 Hz.

这是 1000 条通过自动检查的 G1 仿真运动数据，包含慢走、走路、跑步、前后移动、左右侧移和不同角度的左右转向。每条保存实测与参考两份轨迹。metadata 提供可浏览的索引，motions 提供完整数组；archives 保留原始 CSV 与 JSON，解压后还原 samples/<编号>/ 目录。

These are simulation outputs, not recordings from a physical robot. The collection uses the published pretrained models; it does not train a new planner or controller.

Contents

Motion	Accepted clips	Random command speed bounds
slow_walk	334	0.2–0.8 m/s
walk	333	0.4–1.2 m/s
run	333	1.5–4.0 m/s

• Translation directions: forward, backward, strafe_left, strafe_right, relative to the initial commanded facing direction of the clip.
• Turns: left or right, with command angles 30°, 45°, 60°, 90°, 120°, 135°, 150°, 180°. Turning clips move forward while changing heading over ten seconds.
• Each mode has 20 movement/turn conditions, with 16 or 17 accepted clips per condition.
• Speeds are randomly sampled and rounded to three decimal places. Rejected clips are retried, possibly with a newly sampled speed. Accepted speeds are therefore not guaranteed to follow a uniform distribution.
• Command speed and commanded turn angle are conditions, not guarantees of realized motion. Measured speeds and yaw changes are provided separately.

Quick start

pip install datasets numpy huggingface_hub

from datasets import load_dataset
import numpy as np

# Small metadata table, also used by the default Hub viewer.
index = load_dataset("JianuoCao/Moving", "metadata", split="train")

# One row per complete motion, with both trajectories.
motions = load_dataset("JianuoCao/Moving", "motions", split="train")
clip = motions[0]
t = np.asarray(clip["timestamps_s"], dtype=np.float64)      # [T]
actual = np.asarray(clip["dynamic_qpos"], dtype=np.float64) # [T, 36]
target = np.asarray(clip["kinematic_qpos"], dtype=np.float64)
joint_angles = actual[:, 7:]                               # [T, 29], radians
print(clip["sample_id"], clip["mode"], t.shape, actual.shape)

The train split contains all 1,000 clips. No validation/test split or independent benchmark partition is claimed. Samples may share controller history within a simulation worker, so a random row split should not be assumed to provide independent evaluation.

Repository layout

README.md
data/train-00000-of-00010.parquet       # 10 shards, 100 complete motions each
metadata/metadata.parquet              # 1,000-row index, no large motion arrays
metadata/metadata.jsonl                # same index in a portable text format
metadata/statistics.json               # counts, rates, sampling bounds
metadata/provenance.json               # pipeline and model provenance
metadata/original_file_sha256.json     # checksums for all 5,000 original files
archives/samples-00000-of-00010.tar.gz  # 10 archives, 100 samples each
docs/schema.json                      # fields, units and qpos column order
docs/joint_order.json                  # mapping joint_0 ... joint_28
examples/load_motion.py
previews/ten_motions.mp4               # 10 measured trajectories, 25 fps video
previews/ten_motions.png
previews/paths.png
previews/samples.json
SHA256SUMS                            # checksums of the published export files

Every archive extracts to this structure:

samples/0000/
  trajectory_dynamic.csv
  trajectory_kinematic.csv
  recipe.json
  annotation.json
  quality.json

Archive members are regular files, not symlinks. Their bytes match the original accepted files. Failed attempts, debug logs, machine paths and service response files are not included.

Download and restore the original sample folders:

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="JianuoCao/Moving", repo_type="dataset",
    local_dir="Moving", allow_patterns=["archives/*", "metadata/*", "docs/*", "README.md"]
)

mkdir -p Moving/extracted
for archive in Moving/archives/*.tar.gz; do
  tar -xzf "$archive" -C Moving/extracted
done

The original sample 0000 is then at Moving/extracted/samples/0000/. raw_archive and raw_sample_directory in the index identify the archive and member directory for any sample.

Motion representation and timing

Each raw CSV has 37 columns:

t, base_x, base_y, base_z, qw, qx, qy, qz, joint_0, ..., joint_28

The Parquet representation stores t in timestamps_s and the other 36 columns in each qpos array. Numeric values are stored as float64 and verified to round-trip exactly from the parsed source CSV values. Original CSV bytes are additionally preserved in the archives.

Component	qpos indices	Convention
Base translation	0:3	meters; MuJoCo Z-up coordinates
Base orientation	3:7	unit quaternion, WXYZ
Joint positions	7:36	29 angles in radians; see docs/joint_order.json

Joint groups are left leg (6), right leg (6), waist (3), left arm (7) and right arm (7). Finger joints are not recorded. CSVs do not contain joint velocities, torques, actions, foot-contact labels, images or rewards.

Important coordinate and timing details:

• dynamic_qpos is the measured simulated trajectory. Its translation is in the simulation world coordinates and is not reset to zero for every clip.
• kinematic_qpos is the planner reference recorded by the controller. Its initial XY translation has been subtracted per clip. Its Z value and orientation are retained. Align starting positions before comparing translations; do not assume the two raw XY origins coincide.
• Both sequences share the same per-frame timestamp array, derived from the recording clock. Samples are nominally 50 Hz but have slight timestamp jitter. Use timestamps_s, not a hard-coded timestep, for derivatives.
• The released clips contain 499–501 frames (500,199 frames in total per trajectory kind), with measured rates of approximately 49.913–50.058 Hz. With 500 samples at 50 Hz, the first-to-last timestamp span is approximately 9.98 seconds, while the recorded interval is ten seconds.
• The preview MP4 is resampled to 25 fps for playback. This does not change the source motion frequency.
• Rollouts run sequentially within each of eight isolated workers, with idle settling between clips. Worker restarts occur during failure recovery. World position and controller history can carry across clips within a worker.

Metadata

sample_id is a zero-padded string (0000–0999). mode, direction, turn_direction, turn_angle_deg and command_speed_mps describe the accepted attempt, including any resampled speed. Turn angle is a positive magnitude; turn_direction provides its sign. For straight/sideways clips, the turn fields are null.

dynamic_speed_mps and reference_speed_mps are the mean horizontal path speeds after the first two seconds. Positions are interpolated on a 0.2-second grid and interval distances are divided by 0.2 seconds. This handles curved trajectories; it is not net start-to-end displacement divided by duration.

dynamic_yaw_change_deg is the unwrapped measured yaw change. num_frames, recorded_span_s and measured_fps describe the actual recorded data. accepted_attempt identifies which attempt was accepted, without including the rejected motion files.

text and text_variants contain template-generated English descriptions from annotation.json. They describe commanded behavior and are not human annotations or frame-level semantic labels. Original text variants may contain similar wording and duplicates.

Automatic quality checks

All published clips pass the collection checks. These are heuristic automated checks, not exhaustive physical validation or individual human review.

• At least 400 frames, 37 finite CSV columns, strictly increasing time, and no recording gap above 0.2 seconds.
• Recorded timestamp span between 9.5 and 10.3 seconds; quaternion norm error at most 0.02.
• Measured base height at least 0.35 m and tilt at most 60°.
• Mean measured locomotion speed after two seconds at least 0.08 m/s.
• Straight/sideways motion: displacement-direction cosine at least 0.6 and absolute unintended yaw change at most 35°.
• Turning motion: absolute measured yaw error at most max(25°, 0.3 × commanded angle).

The collector completed 1,116 attempts to obtain 1,000 accepted clips, an attempt pass rate of 89.61%. The first-attempt pass rate was 91.3%. Failed attempts are excluded from the published motion contents.

Mode	Accepted	Attempts	Attempt pass rate
slow_walk	334	345	96.81%
walk	333	354	94.07%
run	333	417	79.86%

Filtering changes the distribution of the accepted data. These trajectories should not be interpreted as evidence that all combinations of commanded speeds and turns can be executed reliably.

Preview

Download the ten-motion overview video.

The videos replay recorded poses, with following cameras and fingers at the model defaults. The path plots align starts and use separate axis scales per panel.

Provenance

Generated using CLAW and NVIDIA GR00T-WholeBodyControl / SONIC. Repository commits, model filenames and collection changes are recorded in metadata/provenance.json.

This repository contains generated motion data and previews, not the upstream model weights, robot assets or source datasets. No additional data license is assigned by this export.