license: cc-by-4.0
task_categories:
- robotics
tags:
- robotics
- tactile
- manipulation
- multimodal
- gelsight
- realsense
- motion-capture
- world-model
- human-collected
- lerobot
pretty_name: React (Tactile-Visual Manipulation)
size_categories:
- 100K<n<1M
configs:
- config_name: motherboard
data_files:
- split: train
path: data/motherboard/meta/**/*.parquet
- config_name: pushT
data_files:
- split: train
path: data/pushT/meta/**/*.parquet
- config_name: rope
data_files:
- split: train
path: data/rope/meta/**/*.parquet
- config_name: validation
data_files:
- split: train
path: data/validation/meta/**/*.parquet
- config_name: all
default: true
data_files:
- split: train
path: data/**/meta/**/*.parquet
React — Multi-Task Tactile-Visual Manipulation
Dense, contact-rich, synchronized multimodal interaction data collected from human hands holding handheld GelSight tactile sensors (no robot arm). Intended for tactile-visual dynamics / world-model learning.
200 min · 361 k frames @ 30 Hz · 3× RGB + 2× GelSight + 2× wrist + OptiTrack · 3 tasks
Per-task scale, contact-force distribution, bimanual contact occupancy, and tactile validity for the wrist-camera era. Panel B shows the ECDF of unsaturated contact force; the steps are real — the force calibration ends in an isotonic (piecewise-constant) stage, so its output is quantised. See Known limits.
What is published
Episodes are cut into clean segments before release: every span a defect
detector flagged is removed, and each remaining span of at least 30 s is
published as its own episode. So the flagged frames are the gaps between
published episodes, and you do not need to consult bad_frames.json before
training on data/<task>/.
| Task | Source recordings | Published segments | Duration | Median contact force |
|---|---|---|---|---|
| motherboard | 12 | 25 | 93.9 min | 2.94 N |
| pushT | 20 | 55 | 59.5 min | 2.94 N |
| rope | 10 | 27 | 47.1 min | 1.38 N |
| (total) | 42 | 107 | 200.5 min |
Two of those medians are the same number because they are the same output value: the force calibration ends in an isotonic regression, so it emits a few thousand discrete levels and 2.94 N is a heavily populated one. See Known limits.
data/validation/ holds a separate 9-segment / 25.2 min set from an
earlier session (2026-09-09) that carries different wrist cameras and a
different calibration epoch. It is not a random held-out split — training on
data/<task> and evaluating there measures domain shift. See its own README.
old_data/ is a separate tree at the repo root, not part of data/. It
holds the pre-2026-09 sessions (motherboard 2026-05, pushT 2026-06), which
have no wrist camera, use an earlier calibration epoch, and — unlike
everything under data/ — are not cut to their clean spans, so a reader
who ignores the bad_frames.json beside them will train on flagged frames.
It is 36.5 GB, most of it depth; see old_data/README.md
and the download table below before cloning the repo whole.
Known limits
Read these before using the force channel or the tactile stream.
Force saturates at 7.87 N. The calibration was fitted on presses up to 8 N, and its final stage is an isotonic regression, which cannot extrapolate — anything above the fitted range is clipped to 7.87 N. Affected fraction of contact frames:
motherboard— 8.3 %pushT— 6.8 %rope— 0.6 %validation— 2.6 %
Treat those frames as right-censored, not as measurements. The force labels also carry roughly 1 N of held-out error, so do not compare models at 0.1 N resolution.
Tactile frames repeat. The GelSight Mini tops out near 17.8 Hz while rows
are written at 29.8 Hz, so only about 54–57 % of rows are a fresh sensor
reading. tactile_{left,right}_is_new marks which. Train tactile dynamics on
the flagged rows, not on all of them.
duration_s is n_frames / 30, the nominal write tick. The RealSense
streams actually run at 29.80 Hz.
Format — LeRobot-style video release
Each episode ships as 7 MP4 video streams (640×480, H.264) + a per-frame parquet of poses and contact metrics. This matches how LeRobot / DROID / Open X-Embodiment ship manipulation data: tiny on disk (data/ is 16.8 GB without depth, against ~1.9 TB of raw recordings), random-access decodable, training-ready.
data/<task>/
├── calibration/ # OptiTrack→camera extrinsics for this task
│ ├── T_mocap_to_cam_{left,middle,right}.json
│ ├── T_gel_to_rigid_{left,right}.json
│ └── calibration.json # epoch, applies-to dates, RMSE, chain
├── videos/<date>/episode_NNN/
│ ├── view_left.mp4 view_middle.mp4 view_right.mp4 # 640×480 RGB
│ ├── tactile_left.mp4 tactile_right.mp4 # 640×480 GelSight
│ └── wrist_left.mp4 wrist_right.mp4 # 640×480 wrist cams
├── meta/<date>/episode_NNN.parquet # one row per frame (see below)
├── episodes.jsonl # one row per episode
├── segments.json # clean-segment index (no bad frames)
├── bad_frames.json # quality intervals per episode
└── previews/<date>/episode_NNN.mp4 # 1280×480 viewer-layout preview
parquet columns (per frame, aligned to video frame i)
| Column | Type | Meaning |
|---|---|---|
frame_idx / frame_index |
int | 0…T-1, matches MP4 frame index |
episode / episode_index |
str / int | source episode key and its 0-based index within the task |
task / task_index |
str / int | task name and index (0=motherboard, 1=pushT, 2=rope) |
force_{L,R}_normal_n |
float32 | estimated normal contact force (N) — censored at 7.87 N, see Known limits |
force_{L,R}_penetration_mm |
float32 | gel penetration depth (mm) |
force_{L,R}_source_frame |
int | the raw tactile frame this force was computed from |
tactile_{L,R}_is_new |
bool | True when the row is a fresh sensor reading, not a repeat |
timestamp |
float64 | camera clock (s) |
sensor_left_pose, sensor_right_pose |
list[7] | OptiTrack world pose of each GelSight (xyz metres + quat xyzw, scalar-LAST — scipy...Rotation.from_quat takes it as-is) |
object_pose |
list[7] | OptiTrack world pose of the manipulated object (NaN where the object body was not tracked — e.g. all pushT) |
tactile_{L,R}_{intensity,area,mixed} |
float32 | contact metrics (computed at full 640×480) |
source_h5_frame |
int | index into the original recording |
Decoded frames are RGB (standard decoder convention) for all five RGB streams.
depth (optional, data/<task>/depth/)
Per-camera depth is shipped as lossless FFV1 16-bit video (gray16le):
data/<task>/depth/<date>/episode_NNN/depth_{left,middle,right}.mkv
- uint16, millimeters;
0= no return / invalid. - Frame
ialigns to the RGB video frameiand parquet rowi. - Decode with PyAV (
frame.to_ndarray()→(480, 640)uint16). cv2 cannot read 16-bit video. - Load via
ReactVideoDataset(..., load_depth=True).
Tasks and calibration
| Task | Dates | Source recordings | Published segments | Duration | Calibration epoch |
|---|---|---|---|---|---|
| motherboard | 2026-09-11, 2026-09-12 | 12 | 25 | 93.9 min | 2026-09-09 |
| pushT | 2026-09-10, 2026-09-11, 2026-09-12 | 20 | 55 | 59.5 min | 2026-09-09 |
| rope | 2026-09-11 | 10 | 27 | 47.1 min | 2026-09-09 |
| validation | 2026-09-09 | 6 † | 9 | 25.2 min | 2026-09-09 |
† validation holds two sessions of the same date and the recorder reuses
episode numbers within a date, so its source_recording strings collide —
four distinct names for six recordings. The count above comes from its own
README, not from the metadata, which cannot express it.
All currently published sessions share the 2026-09-09 extrinsics. Camera extrinsics are used only for the projection overlay; stored poses are OptiTrack world-frame and independent of calibration. Poses are Z-up.
data/validation/ mixes two tasks and two sessions of the same date, and its
episode numbers are not the recorder's — the recorder reuses numbers within
a date, so every row there carries source_recording. Read its README before
using it.
See tasks.json for the machine-readable registry.
Downloading — the whole repo is 57.5 GB; you almost certainly want less
Two things make the full clone much larger than the training data: depth
(36.1 GB, lossless 16-bit) and old_data/ (36.5 GB, of which 31.9 GB
is depth). Both are opt-in subtrees, and skipping them is one argument.
| What you ask for | Size |
|---|---|
data/ only, no depth — the training data |
16.8 GB |
data/ only, with depth |
21.0 GB |
| everything except depth | 21.4 GB |
old_data/ only, no depth |
4.6 GB |
| everything | 57.5 GB |
from huggingface_hub import snapshot_download
# The training data — RGB + tactile + wrist + parquet, no depth (16.8 GB)
snapshot_download("yxma/React", repo_type="dataset",
allow_patterns=["data/*"], ignore_patterns=["*/depth/*"])
# One task (motherboard is the largest, at 9.5 GB)
snapshot_download("yxma/React", repo_type="dataset",
allow_patterns=["data/motherboard/*"],
ignore_patterns=["*/depth/*"])
# Everything, including old_data/ and 36 GB of depth (57.5 GB)
snapshot_download("yxma/React", repo_type="dataset")
Per task, without depth: motherboard 9.5 GB, pushT 3.1 GB, rope 2.7 GB, validation 1.4 GB.
Depth is 16-bit lossless FFV1 and exists for two sessions only:
data/validation/depth/ (4.2 GB) and old_data/*/depth/ (31.9 GB). No
2026-09-10 or later session has depth.
Or use the helper: python examples/download.py --no-depth (see examples/download.py).
The ReactVideoDataset loader never touches depth unless you pass load_depth=True, so depth-free training requires no depth download.
Loading
from examples.react_video_dataset import ReactVideoDataset
ds = ReactVideoDataset("data/motherboard", window_length=16, mode="segment")
sample = ds[0]
# sample["view_middle"]: (16, 480, 640, 3) uint8 RGB
# sample["tactile_left"]: (16, 480, 640, 3) uint8 RGB
# sample["sensor_left_pose"]: (16, 7) float32
mode="segment" iterates clean spans (no bad frames by construction); mode="window" slides over whole episodes and skips bad_frames.json intervals. Backend: PyAV (install decord for faster random access).
World frame — Z-up as published, Y-up on request
Every pose in this release is Z-up: sensor_{left,right}_pose,
object_pose, force_{left,right}_target_pose, and the T_mocap_to_cam
extrinsics in each calibration/. Each subset says so itself, in
episodes.jsonl ("up_axis": "z") and in every calibration file, so nothing
has to infer it from a directory name.
OptiTrack records Y-up. The release is rotated once, by R_x(-90):
(x, y, z) -> (x, -z, y). If you want the raw convention, ask the loader:
ds = ReactVideoDataset("data/motherboard", up_axis="y") # default is "z"
cal = ds.calibration() # comes back in the SAME convention
Take the calibration from the dataset, not from disk. This is a rotation of
the world frame, so poses and T_mocap_to_cam have to move together — one
without the other leaves every projection up to 165 px from the sensor and
raises nothing. react_toolbox.frames is the same conversion if you need it
directly: convert_poses(poses7, to_zup=), as_up_axis(cal, want).
The invariant worth knowing: rotating both changes no picture. Measured on
data/validation over 450 (frame, camera, sensor) combinations, the projected
GelSight centre moves 0.000000 px between the two conventions. Newtons,
timestamps and every video frame are identical either way.
Tactile acquisition latency — does NOT affect any published episode
Recordings up to and including 2026-06-18 had a GelSight-vs-camera capture
lag of ≈15 frames (~0.5 s), caused by a recording-side cv2.VideoCapture V4L2
buffer that was never flushed. It was fixed in the rig on 2026-06-27.
Every episode currently published was recorded on 2026-09-09 or later, and
each one carries per-sensor GelSight timestamps, so the tactile stream is
resampled onto the camera clock during the build. Verified across all 80
published force records: tactile_timestamped is True for every one, and
tactile_align.gel_lag_frames returns 0.
Do not apply a 15-frame shift. Doing so would introduce the half-second misalignment this note used to warn about. The correction is relevant only if you obtain one of the withdrawn 2026-05/06 releases.
Data quality
Per-task bad_frames.json flags intensity_spikes, pose_teleports_{L,R}, ot_loss_{L,R} (OptiTrack track loss). Overall flagged: motherboard 0.90 %, pushT 0.67 %. segments.json already excludes them.
Notes
- Depth is available in the source recordings and will be added under
data/<task>/depth/in a later upload. - One pushT source recording (
episode_004) was corrupt and excluded. - The previous single-task
.ptrelease (episodes/,segments/) is superseded by this video format.