dancher00/go2_object_approach_v1
Go2 Object Approach v1 Keyboard-teleoperated Unitree Go2 EDU trajectories for the task: "Approach the target object and stop in a manipulation-ready pose." The collection contains 202 episodes and 48,889 frames at a nominal 20 Hz (2,444.45 seconds, approximately 40.74 minutes). Episode lengths range from 109 to 502 frames (5.45–25.10 seconds). Data is stored in LeRobot Dataset v3 with Parquet telemetry and H.264 front-camera video. No audio or depth is included. This is a… See the full description on the dataset page: https://huggingface.co/datasets/dancher00/go2_object_approach_v1.
Go2 Object Approach v1
Keyboard-teleoperated Unitree Go2 EDU trajectories for the task: "Approach the target object and stop in a manipulation-ready pose."
The collection contains 202 episodes and 48,889 frames at a nominal 20 Hz (2,444.45 seconds, approximately 40.74 minutes). Episode lengths range from 109 to 502 frames (5.45–25.10 seconds). Data is stored in LeRobot Dataset v3 with Parquet telemetry and H.264 front-camera video. No audio or depth is included.
This is a Go2-only approach dataset. It does not contain arm actions, wrist images, or pick-and-place demonstrations. The associated code repository also provides an experimental Go2 + SO-101 bridge: https://github.com/dancher00/lerobot-unitree-go2
Hardware and collection
Go2 EDU was controlled using keyboard velocity commands. An onboard RealSense D435i provided RGB images; base state came from Unitree SDK2 SportModeState. The collection uses a front camera facing a tabletop target in an indoor workspace. Sampled images show a cylinder and a cup on the table.
Images and telemetry were transported to a laptop over SSH. Camera acquisition was requested at 30 FPS, while the dataset uses a 20 Hz grid and the latest available image. JPEG was used for network transport and H.264 for stored video. Dataset timestamps are nominal frame indices divided by 20, not camera exposure timestamps. Raw session logs are not distributed in this release.
Features
Recorded action ranges are ±0.30 m/s for vx/vy and ±0.60 rad/s for yaw speed. Approximately 52.55% of frames have all three action components equal to zero. These are recorded zero commands; they do not establish that measured robot velocity is zero.
Loading
Load the published dataset from Hugging Face:
from lerobot.datasets.lerobot_dataset import LeRobotDataset
dataset = LeRobotDataset("dancher00/go2_object_approach_v1", video_backend="pyav")
sample = dataset[0]
print(sample["observation.state"], sample["action"])For a local copy, pass root="/path/to/go2_object_approach_v1". Validation used LeRobot 0.6.2 at upstream commit b6ec0060779550c0a157ae34feb89e0cf86012a8.
Validation and limitations
Release preparation on 2026-09-16 checked every table row for valid indices, finite state/actions and feature dimensions; contiguous episode ranges and nominal timestamps; every video's full decode; and video timestamp coverage for all recorded observations. All 28 videos decoded, totaling 48,889 frames. The LeRobot loader also decoded the first, middle and last frame of every episode (606 samples). This verifies structural readability, not task success.
There are no per-episode success labels, rewards, or independent evaluations of policy performance. The supplied split is train: 0:202; no held-out test set is provided. The collection represents one recording setup and should not be treated as evidence of generalization. Repeated pauses/zero commands and camera motion may affect training. Inspect trajectories and define an appropriate evaluation split before reporting results.
The validation can be repeated from the code checkout:
python scripts/validate_dataset.py /path/to/go2_object_approach_v1License and attribution
Dataset: Creative Commons Attribution 4.0 International (CC BY 4.0). Credit dancher00, Go2 Object Approach v1 (2026) and link to the dataset when redistributing or using it in published work. Code in the associated repository is separately licensed under Apache-2.0.