Datasets:

yxma
/

React

license: cc-by-4.0
task_categories:
  - robotics
tags:
  - robotics
  - tactile
  - manipulation
  - multimodal
  - gelsight
  - realsense
  - motion-capture
  - dynamics
  - world-model
  - human-collected
pretty_name: React (Tactile-Visual Manipulation)
size_categories:
  - 100K<n<1M
configs:
  - config_name: episode_metadata
    data_files:
      - split: train
        path: metadata/episodes.parquet
  - config_name: motherboard
    data_files:
      - split: train
        path: episodes/motherboard/**/episode_*.pt
  - config_name: motherboard_segments
    data_files:
      - split: train
        path: segments/motherboard/**/episode_*.segment_*.pt
  - config_name: all
    data_files:
      - split: train
        path: episodes/**/episode_*.pt

React

Dense, contact-rich, synchronized multimodal interaction data collected from human hands holding handheld GelSight tactile sensors — no robot arm involved. Intended for tactile-visual dynamics / world-model learning, not a policy / demonstration dataset.

106 min of robot-free human-hand multimodal interaction · 190,231 frames @ 30 Hz across 3 × RGB-D + 2 × GelSight + 3-body OptiTrack

What's different about this dataset


Robot-arm-free	Recorded directly from a human operator holding two GelSight Mini sensors. No robot kinematics, no embodiment bias, no robot occluding the scene.
Tactile + RGB-D + mocap, simultaneous	Most manipulation datasets ship one of these. React ships all three, synchronized to a common 30 Hz clock.
Contact-dense	64 % of post-trim frames have confirmed tactile contact on at least one sensor — see `figures/contact_intensity_full.png`.
Long, continuous interaction	Recordings are minutes long, not seconds. Median recording duration is 4 min; longest 19 min. Good for short-window sampling of dynamics, not for action-conditioned policy learning.

At a glance


Embodiment	Human hands (no robot) — handheld GelSight sensors with motion-capture rigid bodies
Intended use	Dynamics / world-model learning over short multimodal windows. Sample short trajectories (1 s – 10 s); recording-file boundaries are not action boundaries.
Total synchronized duration	105.7 min at 30 Hz (190,231 multimodal frames, post-trim)
Bimanual tactile-contact time	64.3 % of post-trim frames (3,302 contact events, median 0.73 s; see `figures/dataset_figures/F2_contact_event_duration_histogram.png` and `metadata/episodes.parquet` for per-file numbers)
Cameras	3× Intel RealSense D415 (color + depth), 480×640, 30 FPS
Tactile	2× GelSight Mini (left, right), handheld
Motion capture	OptiTrack VRPN, 3 rigid bodies, ~120 Hz
Tasks	`motherboard` (more coming)
License	CC-BY-4.0

Recording sessions

Date	Kind	Active sensors	Notes
2026-05-10	session	left + right	First full bimanual session.
2026-05-11	session	left + right	Largest session. A handful of GelSight LED-flicker frames + one mocap teleport; see `bad_frames.json`.

| 2026-05-19 | session | left + right | New session, multi-cam (view_left/middle/right) end-to-end. Curation via reproducible detect_bad_intervals.py ruleset (see docs/curation_pipeline.md). |

See tasks.json for the machine-readable registry (per-date active_sensors, etc.).

OT-uninitialized prefixes trimmed. Three episodes had OptiTrack offline at the start of recording (1–11 min each); those prefixes have been cut from the published .pt files (_contact_meta.trim_offset per file). Future recordings use an OT watchdog that refuses to start an episode unless mocap is streaming. Full story: docs/caveats.md.

Data quality

Mode	Frames	%	Files	Cause
GelSight LED flicker	56	0.029 %	5	Single-frame LED dropout, recovers next frame
OptiTrack pose teleport	56	0.029 %	3	Solver flip (translation > 5 m/s or angular > 15 rad/s)
OptiTrack track loss	1,680	0.883 %	6	Marker briefly left mocap-volume / camera FOV mid-episode
Total (union)	1,768	0.929 %	11

Every flagged interval is in bad_frames.json keyed by episode/episode_* with TRIMMED-pt frame indices. A richer per-event view (with cross-modal motion + OT-gap + angular-velocity stats) lives in freeze_intervals.json. Skip-list usage is shown below and in docs/quality.md. Long start-of-episode OT-uninitialized prefixes (the dominant problem in the raw recordings) have already been trimmed from the published .pt files — see docs/caveats.md.

Two layouts: `episodes/` vs `segments/`

The same recordings are shipped two ways depending on what your code wants to do:

episodes/<task>/<date>/episode_*.pt — one file per recording. Includes bad intervals (LED flicker, pose teleport, OT track loss) inside; downstream code is expected to filter them out using bad_frames.json. Each file carries all three RealSense views (view_left, view_middle, view_right) plus both GelSights.
segments/<task>/<date>/episode_*.segment_*.pt — same recordings, but pre-sliced into contiguous clean segments at every bad-frames boundary. No bad_frames.json lookup needed; the data is clean by construction. Index lookup via segments.json. Each segment's _contact_meta.source_h5_frame_range maps it back to the original recording. The example ReactSegmentDataset (examples/react_segment_dataset.py) consumes these.

Both layouts have identical content (same source recordings, same frame data); only the file boundaries differ.

Quick start

# Load by task with `datasets`
from datasets import load_dataset
ds = load_dataset("yxma/React", "motherboard", split="train")

Or grab a single recording file directly:

import torch
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="yxma/React", repo_type="dataset",
    filename="episodes/motherboard/2026-05-11/episode_003.pt",
)
ep = torch.load(path, weights_only=False)
# ep["view"]                                 (T, 3, 128, 128) uint8 — overhead cam
# ep["tactile_left"], ep["tactile_right"]    (T, 3, 128, 128) uint8
# ep["sensor_left_pose"], ep["sensor_right_pose"]
#                                            (T, 7) float32 — xyz + quaternion
# ep["timestamps"]                           (T,)  float64
# Plus per-frame contact metrics:            tactile_{side}_{intensity, area, mixed}

Sampling short windows for dynamics learning — drop windows that overlap any flagged interval:

import json

with open("bad_frames.json") as f:
    bad = json.load(f)["episodes"]   # frame indices are TRIMMED-pt coordinates

def is_clean_window(episode_key, t_start, t_end):
    """True iff [t_start, t_end] doesn't intersect any flagged span."""
    bf = bad[episode_key]
    intervals = (bf["intensity_spikes"]
                 + bf["pose_teleports_L"] + bf["pose_teleports_R"]
                 + bf["ot_loss_L"] + bf["ot_loss_R"])
    return all(not (s <= t_end and e >= t_start) for s, e in intervals)

Currently 1,768 / 190,231 frames (0.93 %) are flagged across 11 of 27 files — see docs/quality.md for the per-mode breakdown and more filtering recipes. The example dataloader below does this filtering for you when skip_bad_frames=True.

Example dataloader — short contact-rich windows

A reference PyTorch Dataset is shipped under examples/react_window_dataset.py. It scans the processed .pt files, applies the contact filter, drops windows that overlap bad_frames.json, and respects the per-date active_sensors field from tasks.json.

from examples.react_window_dataset import ReactWindowDataset
from torch.utils.data import DataLoader

ds = ReactWindowDataset(
    data_root="episodes/motherboard",
    bad_frames_path="bad_frames.json",
    tasks_json_path="tasks.json",
    window_length=16,            # frames per window
    stride=1,                    # within-window stride (1 = consecutive)
    window_step=16,              # step between window starts (overlap control)
    contact_metric="mixed",      # "intensity" | "area" | "mixed"
    tactile_threshold=0.4,
    min_contact_fraction=0.6,    # ≥ 60 % of window frames must have contact
    which_sensors="any",         # "any" | "both" | "left" | "right"
    skip_bad_frames=True,
    respect_active_sensors=True,
)
print(len(ds), "windows")
loader = DataLoader(ds, batch_size=8, shuffle=True, num_workers=2)

With the defaults shown above, the dataset assembles ~9.2 k contact-rich 16-frame windows across the 27 recordings. Each sample is a dict of (T, …) tensors plus metadata (episode, frame_start, active_sensors, …).

Example output

Four random windows, time runs left→right; each cell is view | tactile_left | tactile_right with sensor frame axes (X red, Y green, Z blue-ish) projected onto the view:

One window played frame-by-frame with the sensor-frame overlay:

Full demo script: examples/demo_react_window.py.

Recording-file previews

Per-episode previews live under figures/episode_previews/ as inline-renderable MP4s. Browse all 32 episodes (collapsed by default) on — click any row to preview that episode inline. Each shows 60 frames evenly sampled across the episode in the recording-viewer layout: 3 RealSense cameras with projected GelSight axes, GelSight raw + diff thumbs, OptiTrack pose text panel. (The on-disk recording unit is called an "episode" purely for file naming — these boundaries don't carry semantic / action meaning for this dataset.)

Repository layout

README.md                                        # this file
tasks.json                                       # task / session registry
bad_frames.json                                  # data-quality skip-list
episodes/<task>/<date>/episode_*.pt    # per-file tensors
figures/                                         # previews + analysis figures
docs/                                            # extended documentation

File	Contents
`docs/recording.md`	Hardware setup, camera serials, sensor + mocap layout, robot-free collection method
`docs/schema.md`	Full `.pt` field reference and contact-metric definitions
`docs/quality.md`	Data-quality breakdown (per-mode), `bad_frames.json` schema, dataloader recipe, inspection figures
`docs/figures.md`	Dataset statistics + analysis gallery (F1–F8)
`docs/caveats.md`	Known caveats and roadmap

License

Released under Creative Commons Attribution 4.0 (CC-BY-4.0).

Citation

If you use this dataset, please cite (TODO: add bibtex).