TacVerse · XTac-UMI-G1 Head-view
Collection
Bimanual visuo-tactile demos with XTac-UMI-G1 grippers and Pico headset stereo cameras. LeRobot v3.0 format. • 14 items • Updated • 1
896 episodes · 30 fps · 8 cameras · 640×480 h264
Representative frames from TacVerse's bimanual demonstrations.
Collected with XTac-UMI-G1 grippers, released as LeRobot datasets.
This dataset was created using LeRobot.
Access
This dataset is public. Download it and load it with LeRobot:
hf download TacVerse/xtac-umi-g1-remove-oring-a --repo-type dataset --local-dir ./xtac-umi-g1-remove-oring-a
from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("TacVerse/xtac-umi-g1-remove-oring-a", root="./xtac-umi-g1-remove-oring-a")
Privacy
Faces of people appearing in the camera streams (wrist and head cameras) have been anonymized: they were detected automatically and pixelated before release. The tactile streams contain no people and are unchanged. Automatic detection can miss a face; if you find one that is not masked, please contact us via the TacVerse landing page and we will fix it.
Annotations
Every episode carries automatic language annotations produced by a vision-language model from the head-camera
and tactile frames. They are model-generated descriptions, not human labels; spot-check them before relying on them.
The frame data and meta/info.json are unchanged, so the dataset loads the same way with any LeRobot version.
meta/language_persistent.parquet— one row per language atom, using the row schema of LeRobot'slanguage_persistentcolumn (role,content,style,timestamp,camera,tool_calls) plusepisode_index.style="task_aug"is a one-sentence summary of what the episode actually shows;style="subtask"rows are the step-by-step segmentation. A subtask becomes active at itstimestamp(seconds from episode start) and stays active until the next subtask row, so exactly one subtask is active at any frame. Steps done by one gripper are marked(left arm)/(right arm); steps the two grippers start together are joined asleft: …; right: ….meta/annotations.parquet— one row per episode with the full annotation as JSON (labels_json): scene objects, the raw timeline (per gripper, with object / destination / progress), key events, state changes, a scene graph and a completion judgement, plus the instruction the model was given (annotated_instruction) and themodel.
Active subtask for each frame:
import pandas as pd
from lerobot.datasets.lerobot_dataset import LeRobotDataset
root = "./xtac-umi-g1-remove-oring-a"
ds = LeRobotDataset("TacVerse/xtac-umi-g1-remove-oring-a", root=root)
lang = pd.read_parquet(f"{root}/meta/language_persistent.parquet")
sub = lang[lang["style"] == "subtask"].sort_values(["episode_index", "timestamp"])
def subtask_at(episode_index: int, t: float) -> str:
rows = sub[(sub.episode_index == episode_index) & (sub.timestamp <= t + 1e-6)]
return rows.iloc[-1]["content"]
frame = ds[0]
print(subtask_at(int(frame["episode_index"]), float(frame["timestamp"])))
Dataset Structure
{
"codebase_version": "v3.0",
"robot_type": "xtac_umi_g1",
"collection_stack": "xense-taccap-lerobot",
"total_episodes": 896,
"total_frames": 1729574,
"total_tasks": 1,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 200,
"fps": 30,
"splits": {
"train": "0:896"
},
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path": "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4",
"features": {
"action": {
"dtype": "float32",
"names": [
"left_tcp.x",
"left_tcp.y",
"left_tcp.z",
"left_tcp.r1",
"left_tcp.r2",
"left_tcp.r3",
"left_tcp.r4",
"left_tcp.r5",
"left_tcp.r6",
"left_gripper.pos",
"right_tcp.x",
"right_tcp.y",
"right_tcp.z",
"right_tcp.r1",
"right_tcp.r2",
"right_tcp.r3",
"right_tcp.r4",
"right_tcp.r5",
"right_tcp.r6",
"right_gripper.pos",
"head.x",
"head.y",
"head.z",
"head.r1",
"head.r2",
"head.r3",
"head.r4",
"head.r5",
"head.r6"
],
"shape": [
29
]
},
"observation.state": {
"dtype": "float32",
"names": [
"left_tcp.x",
"left_tcp.y",
"left_tcp.z",
"left_tcp.r1",
"left_tcp.r2",
"left_tcp.r3",
"left_tcp.r4",
"left_tcp.r5",
"left_tcp.r6",
"left_gripper.pos",
"right_tcp.x",
"right_tcp.y",
"right_tcp.z",
"right_tcp.r1",
"right_tcp.r2",
"right_tcp.r3",
"right_tcp.r4",
"right_tcp.r5",
"right_tcp.r6",
"right_gripper.pos",
"head.x",
"head.y",
"head.z",
"head.r1",
"head.r2",
"head.r3",
"head.r4",
"head.r5",
"head.r6"
],
"shape": [
29
]
},
"observation.images.left_tactile_left": {
"dtype": "video",
"shape": [
400,
700,
3
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 400,
"video.width": 700,
"video.codec": "h264",
"video.pix_fmt": "yuv420p",
"video.is_depth_map": false,
"video.fps": 30,
"video.channels": 3,
"has_audio": false
}
},
"observation.images.left_tactile_right": {
"dtype": "video",
"shape": [
400,
700,
3
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 400,
"video.width": 700,
"video.codec": "h264",
"video.pix_fmt": "yuv420p",
"video.is_depth_map": false,
"video.fps": 30,
"video.channels": 3,
"has_audio": false
}
},
"observation.images.left_wrist": {
"dtype": "video",
"shape": [
480,
640,
3
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 480,
"video.width": 640,
"video.codec": "h264",
"video.pix_fmt": "yuv420p",
"video.is_depth_map": false,
"video.fps": 30,
"video.channels": 3,
"has_audio": false
}
},
"observation.images.right_tactile_left": {
"dtype": "video",
"shape": [
400,
700,
3
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 400,
"video.width": 700,
"video.codec": "h264",
"video.pix_fmt": "yuv420p",
"video.is_depth_map": false,
"video.fps": 30,
"video.channels": 3,
"has_audio": false
}
},
"observation.images.right_tactile_right": {
"dtype": "video",
"shape": [
400,
700,
3
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 400,
"video.width": 700,
"video.codec": "h264",
"video.pix_fmt": "yuv420p",
"video.is_depth_map": false,
"video.fps": 30,
"video.channels": 3,
"has_audio": false
}
},
"observation.images.right_wrist": {
"dtype": "video",
"shape": [
480,
640,
3
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 480,
"video.width": 640,
"video.codec": "h264",
"video.pix_fmt": "yuv420p",
"video.is_depth_map": false,
"video.fps": 30,
"video.channels": 3,
"has_audio": false
}
},
"observation.images.head_left": {
"dtype": "video",
"shape": [
480,
640,
3
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 480,
"video.width": 640,
"video.codec": "h264",
"video.pix_fmt": "yuv420p",
"video.is_depth_map": false,
"video.fps": 30,
"video.channels": 3,
"has_audio": false
}
},
"observation.images.head_right": {
"dtype": "video",
"shape": [
480,
640,
3
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 480,
"video.width": 640,
"video.codec": "h264",
"video.pix_fmt": "yuv420p",
"video.is_depth_map": false,
"video.fps": 30,
"video.channels": 3,
"has_audio": false
}
},
"timestamp": {
"dtype": "float32",
"shape": [
1
],
"names": null
},
"frame_index": {
"dtype": "int64",
"shape": [
1
],
"names": null
},
"episode_index": {
"dtype": "int64",
"shape": [
1
],
"names": null
},
"index": {
"dtype": "int64",
"shape": [
1
],
"names": null
},
"task_index": {
"dtype": "int64",
"shape": [
1
],
"names": null
},
"observation.tracker_confidence": {
"dtype": "float32",
"names": [
"left_tracker.confidence",
"right_tracker.confidence"
],
"shape": [
2
]
}
},
"tacflow_stage": "postprocess.anonymize",
"tcp_frame_convention": "mirrored-v1"
}
Sensor key map
Tactile keys map to fingertip sensors; wrist keys to gripper cameras; head keys to the headset's stereo eyes.
Citation
BibTeX:
@misc{tacverse2026,
title = {TacVerse: A Large-Scale Bimanual Visuo-Tactile Manipulation Dataset},
author = {Xense Robotics},
year = {2026},
howpublished = {\url{https://huggingface.co/datasets/TacVerse/TacVerse}},
note = {Collected with the XTac-UMI-G1 handheld gripper and Xense Neutron visuotactile sensors}
}
- Downloads last month
- 631