VatinZ/VT_Hand

DexH13-VT

DexH13-VT is a real-world vision–tactile teleoperation dataset of a right DexH13 hand. Each episode is one grasp of a household object, with joint positions, dense tactile forces, a palm camera, and an external RGB camera, all aligned at 30 Hz.

The release contains 383 episodes, 132,170 frames (1.84 hours), and 49 objects. Episodes are ep_000000 through ep_000382, with no gaps.

On the Hugging Face Hub the dataset is one archive, dexh13_vt.tar.zst, next to this card. Extract it, then the paths below match the unpacked dexh13_vt/ directory:

tar --zstd -xf dexh13_vt.tar.zst

Summary

Episodes	383
Frames	132,170
Duration	6,622 s (mean 17.3 s, range 9.9–48.3 s)
Rate	30 Hz
Objects	49
Hand	DexH13, right, 16 DoF
Operator	one (op_01)
Collected	2026-08-12 to 2026-08-16 (UTC)
Size on disk	about 15 GB
Schema	0.1.0

Tasks

task_type	Episodes
pick_place	140
pick_hold	132
deformable_grasp	99
texture_grasp	12

Grasp labels

grasp_type	Episodes
pinch_2	237
envelop	75
tripod	65
other	6

Object attributes, counted by object rather than by episode: 18 boxes, 11 spheres, 10 cylinders, 10 irregular shapes; 22 hard, 10 medium, 17 soft. Size grades are near-equal bins of the grasp-relevant size size_ref_mm: XS 15.5–36 mm (13), S 38–48.3 mm (12), M 49–61 mm (12), L 62–99 mm (12).

Layout

meta/dataset.json                 tactile layout and coordinate notes
meta/object_labels.json           shape, hardness, and size for each object_id
calibrations/                     camera intrinsics, hand–eye, and tactile baseline
episodes/ep_XXXXXX/
  meta.json                       episode record
  label.json                      operator outcome
  events.json                     start, first contact, release, end
  phases.json                     approach / contact / release / idle
  timeline.npz                    per-frame index, time, and copies of the streams below
  qpos/qpos.npy                   (T, 16) float32, measured joint position, radians
  control/cmd_qpos.npy            (T, 16) float32, commanded joint position, radians
  control/control_mode.npy        (T,) uint8
  tactile/taxel_force.npy         (T, 1140, 3) float32, Newtons, sensor frame
  tactile/tip_force.npy           (T, 11, 3) float32, Newtons, one vector per module
  vision/palm/000000.jpg          640×480 JPEG, palm camera, one file per frame
  vision/external/000000.jpg      640×480 JPEG, RealSense D435i color
  vision/*_timestamps.npy         (T,) float64

T is num_frames in meta.json (179–692, mean 345). JPEG names are the frame index, zero-padded to 6 digits. In every episode, timeline.npz palm_frame_idx and external_frame_idx equal frame_idx, and both image folders contain exactly num_frames files. timeline.npz also stores t_wall, t_mono, tip_force_norm (T, 11), total_force (T,), n_contact_modules (T,), and qpos_clipped (T,) bool.

Joint order is the DexH13 SDK order. Values are radians.

Tactile layout

taxel_force[:, start:end, :] is one sensing region. Order and slices are fixed and also written in meta/dataset.json. Stored values are Newtons (raw readings in units of 0.1 N, multiplied by 0.1), in the sensor-local xyz frame. tip_force follows the same module order.

Index	Module	Taxels	Slice
0	index_base	120	0:120
1	index_mid	60	120:180
2	index_tip	120	180:300
3	middle_base	120	300:420
4	middle_mid	60	420:480
5	middle_tip	120	480:600
6	ring_base	120	600:720
7	ring_mid	60	720:780
8	ring_tip	120	780:900
9	thumb_base	120	900:1020
10	thumb_tip	120	1020:1140

Cameras

Both streams are 640×480 color at 30 Hz. Depth is not recorded.

Camera	Device	Intrinsics
Palm	DexH13 palm camera	calibrations/palm_camera/intrinsics_640x480.json (fx ≈ 352.3, fy ≈ 352.8, cx ≈ 318.6, cy ≈ 219.9), plus distortion and the hand–eye matrix
External	Intel RealSense D435i, color only	calibrations/external_camera/realsense_640x480.json (fx ≈ 607.6, fy ≈ 607.0, cx ≈ 320.2, cy ≈ 239.5)

Labels

label.json is filled by the operator at the end of the episode. Every episode has outcome: success and object_feasible: true.

events.json and phases.json are produced by one automatic contact detector, with the same thresholds on every episode: a module counts as touching when its force exceeds 0.3 N, contact requires at least two modules and 0.8 N total, sustained for 100 ms, and the first 0.5 s of the episode is ignored. The primary phases are approach, contact, release, and idle.

Phase / event	Episodes
approach	383
contact / first_contact	368
release	72
idle (after release)	72

Fifteen episodes never cross the contact threshold, so they have no contact phase.

meta/object_labels.json maps each object_id to shape (sphere, cylinder, box, irregular), hardness (soft, medium, hard), size (XS, S, M, L), size_mm (three measured extents), and size_ref_mm (one grasp-relevant scalar: diameter for a sphere; a diameter-like value for a cylinder; the short side for a box or irregular object, with a wrapped cross-section for very flat packs). size_continuous is size_ref_mm / 99. Forty-seven objects were measured with a ruler. bread and bread_in_a_puffy_bag use a vision-model estimate (size_source: vlm_estimate).

Objects

Episode counts. Grasp and task columns list the labels present for that object.

object_id	Ep.	Shape	Hardness	Size	mm	Grasps	Tasks
wooden_box	18	box	hard	M	50.0	envelop, pinch_2, tripod	pick_hold, pick_place
climbing_hold	15	irregular	hard	XS	33.0	envelop, pinch_2, tripod	pick_hold, pick_place
moon_ball	15	sphere	medium	L	76.0	envelop, pinch_2, tripod	pick_hold, pick_place
tennis	13	sphere	hard	M	61.0	pinch_2, tripod	pick_hold, pick_place
bottle	12	cylinder	medium	L	96.0	envelop, pinch_2, tripod	deformable_grasp, pick_place
umbrella	12	cylinder	medium	S	43.0	envelop, other, pinch_2	pick_hold, pick_place
vertical_striped_paper_tube	12	cylinder	medium	M	54.0	other, pinch_2	pick_hold, pick_place
weighted_bubble_roll	12	cylinder	soft	L	62.0	envelop, pinch_2, tripod	deformable_grasp, pick_place
sponge_block	11	box	soft	L	99.0	envelop, pinch_2, tripod	deformable_grasp, pick_place
sandbag	10	box	soft	L	64.0	pinch_2, tripod	deformable_grasp, pick_place
baseball	9	sphere	hard	L	69.0	pinch_2, tripod	pick_hold, pick_place
bouncy_ball	9	sphere	soft	L	64.0	pinch_2, tripod	pick_hold, pick_place
bubble_wrap_roll	9	cylinder	soft	M	49.0	envelop, pinch_2, tripod	deformable_grasp, pick_place
flower_foam	9	box	medium	L	65.0	pinch_2, tripod	deformable_grasp, pick_place
hard_massage_ball	9	sphere	hard	L	64.0	envelop, pinch_2	pick_place, texture_grasp
iron_ball	9	sphere	hard	M	55.0	pinch_2, tripod	pick_hold, pick_place
metal_box	9	box	hard	XS	31.1	pinch_2, tripod	pick_hold, pick_place
paper_tube	9	cylinder	medium	S	48.3	pinch_2, tripod	pick_hold, pick_place
pillow	9	box	soft	XS	36.0	pinch_2, tripod	deformable_grasp, pick_place
rockery_model	9	irregular	hard	S	48.0	envelop, pinch_2	pick_hold, pick_place
water_filled_bag	9	irregular	soft	S	47.6	envelop, pinch_2, tripod	deformable_grasp
wooden_ball	9	sphere	hard	L	63.0	pinch_2, tripod	pick_hold, pick_place
stiff_slime	7	irregular	medium	M	50.0	pinch_2	deformable_grasp, pick_place
checkered_cylinder	6	cylinder	hard	M	52.0	envelop	pick_place, texture_grasp
green_dragon	6	box	hard	XS	25.0	pinch_2	pick_hold, pick_place
massage_stick_with_dots	6	cylinder	hard	L	70.0	envelop	pick_hold, pick_place
molecular_model	6	irregular	medium	M	53.0	pinch_2	pick_hold, pick_place
one_of_circles	6	box	hard	XS	25.0	pinch_2	pick_hold, pick_place
perforated_plastic_block	6	box	hard	M	50.0	pinch_2	pick_hold, pick_place
pingpong_ball	6	sphere	hard	S	38.0	pinch_2	pick_hold, pick_place
plastic_box	6	box	hard	XS	22.0	pinch_2	pick_hold, pick_place
red_dragon	6	box	hard	XS	25.0	pinch_2	pick_hold, pick_place
resin_block	6	box	hard	S	39.0	pinch_2	pick_hold, pick_place
rubber_ball	6	sphere	medium	S	48.0	pinch_2	deformable_grasp, pick_place
soft_ball	6	sphere	soft	L	63.0	pinch_2	deformable_grasp, pick_place
soft_slime	6	irregular	soft	XS	30.9	envelop, pinch_2	deformable_grasp
spiky_massage_stick	6	cylinder	hard	M	54.0	envelop	pick_hold, pick_place
striped_cylinder	6	cylinder	hard	M	52.0	envelop	pick_place, texture_grasp
white_dragon	6	box	hard	XS	25.0	pinch_2	pick_hold, pick_place
wooden_block	6	box	hard	M	60.0	pinch_2, tripod	pick_hold
yarn_ball	6	sphere	soft	S	45.0	pinch_2	deformable_grasp, pick_place
plasticine	5	irregular	soft	XS	28.0	pinch_2	deformable_grasp, pick_place
silicone_toy	5	irregular	soft	S	48.0	pinch_2	deformable_grasp, pick_place
small_doll	5	irregular	soft	XS	31.0	pinch_2	deformable_grasp, pick_place
bread	3	box	soft	S	40.0	pinch_2	deformable_grasp
bread_in_a_puffy_bag	3	box	soft	S	42.0	pinch_2	deformable_grasp
fluff_filled_resealable_bag	3	box	soft	XS	29.8	pinch_2	deformable_grasp
ice_strip	3	irregular	soft	XS	15.5	pinch_2	deformable_grasp
take_out_box	3	box	medium	S	41.0	pinch_2	deformable_grasp

Loading

import json
from pathlib import Path

import numpy as np
from PIL import Image

root = Path("dexh13_vt")  # or the directory from snapshot_download
episode = root / "episodes" / "ep_000001"
meta = json.loads((episode / "meta.json").read_text())
labels = json.loads((root / "meta" / "object_labels.json").read_text())

qpos = np.load(episode / "qpos" / "qpos.npy")             # (T, 16)
taxel = np.load(episode / "tactile" / "taxel_force.npy")  # (T, 1140, 3)
tip = np.load(episode / "tactile" / "tip_force.npy")      # (T, 11, 3)
frame = Image.open(episode / "vision" / "palm" / "000000.jpg")
obj = labels[meta["object_id"]]

from huggingface_hub import snapshot_download

path = snapshot_download("YOUR_ORG/dexh13-vt", repo_type="dataset")

There is no official train/test split. To test generalization to new objects, hold out entire object_id values rather than random episodes of the same object.

Collection

An operator teleoperated the hand, then switched to a hold with a manual key (control_policy.mode_primary is teleop_then_hold, hold_trigger is manual_G). Human hand motion was retargeted onto the DexH13 (checkpoint tag modellatest) with rectification, wrist recentering, an EMA of 0.35, and thumb equalization. apply_pinch_remap is false on every episode. control/control_mode.npy is present and is 0 on every frame; the policy is recorded in meta.json, not in that array.

qpos_clipped in timeline.npz is true on 15,998 frames (12.1%), meaning at least one joint was at its limit on that frame.

Considerations

• One operator and one right hand. Results do not cover left hands or other people.
• Every episode is labeled a success.
• Contact and release labels are automatic. Fifteen episodes have no detected contact, and a release is marked on only 72 episodes.
• The RealSense recording is color only.
• Palm and external images show the operator's hand and the lab bench. operator_id is a pseudonym (op_01).
• Joint names are not stored. Use the DexH13 SDK joint order.
• bread and bread_in_a_puffy_bag sizes are vision-model estimates, not ruler measurements.

License

Creative Commons Attribution 4.0 International (cc-by-4.0). You may share and adapt the dataset, including for commercial use, if you give appropriate credit.

Citation

Cite this dataset repository when you use the data. A paper citation will be added here when it is available.