Datasets:

yxma
/

React

# get the toolbox (it lives in the dataset repo)
huggingface-cli download yxma/React toolbox/ --repo-type dataset --local-dir react
cd react
pip install numpy pyarrow av opencv-python   # core
pip install torch scipy                       # optional: depth

import react_toolbox as T
from huggingface_hub import hf_hub_download

# load tactile frames + per-frame metadata
vid  = hf_hub_download("yxma/React", "data/motherboard/videos/2026-05-10/episode_000/tactile_left.mp4", repo_type="dataset")
meta = T.load_meta(hf_hub_download("yxma/React", "data/motherboard/meta/2026-05-10/episode_000.parquet", repo_type="dataset"))
frames = T.load_video(vid, range(200))                       # (200, 480, 640, 3) RGB

# 1) reference (no-contact) frame + difference image
ref  = T.get_reference(frames, mode="p01", intensity=meta["tactile_left_intensity"][:200])
diff = T.difference(frames[100], ref)                        # Sparsh-style signed diff

# 2) contact detection (calibration-free)
mask    = T.contact_mask(frames[100], ref)                   # (480,640) bool
metrics = T.contact_metrics(frames[100], ref)               # {intensity, area, mixed} == dataset scalars

# 3) approximate depth / height map (pretrained nnmini, no calibration)
from react_toolbox import depth
h = depth.height_map(frames[100], ref)                       # (480,640) relative height

# 4) visualization (all return RGB uint8)
hm  = T.diff_heatmap(frames[100], ref)
ov  = T.contact_overlay(frames[100], ref)

# 5) camera projection (per-task extrinsics)
cal = T.load_calibration("data/motherboard")                # after snapshot_download of calibration/
uv  = T.project_gel_to_pixel(meta["sensor_left_pose"][100], cal["gel_left"], cal["cams"]["middle"])

# 5b) SEE it land — a coordinate cannot tell you whether it is on the sensor,
#     and every projection defect this dataset has shipped was obvious in a
#     picture and invisible in a number (21-36 px, 35-73 px, 155-223 px, all
#     of them shaped like a slightly miscalibrated rig).
img = T.draw_projection(cam_frame, meta["sensor_left_pose"][100],
                        cal["gel_left"], cal["cams"]["middle"],
                        force_n=meta["force_left_normal_n"][100],
                        target_pose7=meta["force_left_target_pose"][100])

# 5c) or check it without looking: every episode's parquet declares the world
#     frame its poses are in, and carries a projection fingerprint you can
#     recompute. A pose array in the wrong frame misses it by 150+ px.
dec = T.read_world_frame(episode_parquet)     # {"world_frame", "raw_h5_offset_m", "fingerprint"}
err = T.verify_world_frame(meta["sensor_left_pose"], "left", "motherboard", dec)
assert err < 6.0, f"poses are not in the declared frame: {err:.1f} px off"

# 6) derive actions from handheld poses
act   = T.next_state_action(meta["sensor_left_pose"])       # next-frame absolute state
delta = T.delta_pose_action(meta["sensor_left_pose"])       # frame-to-frame delta

Run the full demo (saves a montage):

python -m react_toolbox.demo --with_depth

Functions

Module	Function	Notes
`io`	`load_video`, `load_meta`, `episode_paths`	PyAV / OpenCV decode → RGB
`reference`	`get_reference`, `difference`, `l2_diff`	p01 / first / running-avg reference; Sparsh signed diff
`contact`	`contact_mask`, `contact_metrics`, `contact_centroid`	diff→threshold→largest component; reproduces dataset scalars
`depth`	`height_map`, `normals`, `poisson_integrate`	approximate, uncalibrated — pretrained markerless-Mini net (`nnmini.pt`, fetched on demand) + DCT Poisson
`viz`	`diff_heatmap`, `contact_overlay`, `reference_compare`, `depth_view`, `height_to_pointcloud`	all RGB uint8; `depth_view` is grayscale by default (standard GelSight height map), pass `cmap=` for a colormap
`calibration`	`load_calibration`, `project_gel_to_pixel`	per-task extrinsics (motherboard=May-12, pushT=June-26)
`actions`	`next_state_action`, `delta_pose_action`, `integrate_delta`	handheld pose → IL/world-model targets

Notes & limits

No markers on this sensor → no marker-flow / shear-field utilities (would need a markered gel).
Depth is approximate / relative, not metric: it uses a network pretrained on a reference GelSight Mini (no per-unit calibration). Verified to track contact deformation locally (height elevation under contact correlates 0.74 with contact intensity), but the global height includes the gel's baseline curvature. For metric depth, collect a ball-indenter calibration.
nnmini.pt weights are © GelSight Inc (GPL-3.0); only the weight file is fetched on demand — no GPL code is bundled. The toolbox code itself is MIT.
All decoded frames are RGB. Contact scalars use L2 threshold tau=8.0 (dataset default).