Tidel AI · Tactile Glove + Hand Pose Dataset

What's in the package›
PER EPISODE
- skeleton.mp4cam-b | cam-c side-by-side, hand-skeleton overlay burned in
- depth.mp4stereo depth, false-color
- context.mp4cam-a | cam-d wide context strip
- tactile.jsonper-frame: MANO 21-joint xyz (mm) · per-region force · per-finger bend · IMU roll/pitch/yaw + gyro/accel, both hands
- annotationaction segments + task outcome (EP-01 and EP-05…23)
RAW SOURCE (per episode, not shipped on this page)
- cam-a/b/c/d/rgb/~3650 JPEG frames each, 30 FPS
- cam-b, cam-c /mask/segmentation masks
- glove/jq_glove_ttyACM0,1/glove.csv185 Hz raw taxel + IMU stream, ~22.7k rows
- hand.jsoncamera intrinsics/extrinsics + per-frame MANO pose/shape
Field dictionary›

FORCE

Sum of a region's raw taxel readings. Uncalibrated — arbitrary units, not Newtons, until a per-glove force curve is supplied.

BEND

Per-finger flex sensor angle, degrees.

IMU ROLL/PITCH/YAW

Derived from the glove's onboard quaternion (Euler decomposition), not from vision.

MANO JOINTS

21-keypoint hand pose from the camera-side reconstruction pipeline (hand.json), in mm, camera reference frame. This is a separate estimate from the glove IMU — the two are not fused on this page.
Known limitations · this sample›

GLOVE ↔ HAND PAIRING CONFIDENCE

EP-01…04 (glove rig): left/right assignment is resolved by cross-correlating glove gyro energy against MANO wrist motion. Correlation is weak on every episode (r ≈ 0.05–0.15) — the assignment shown is the better of two options, not a confident match. Treat these left/right labels as provisional.

EP-05…23 (stereo rig): this does not apply. Each glove ships as its own file on a shared hardware timebase, so the assignment is explicit rather than inferred. Camera↔IMU alignment measured across all 19 captures stays within ±1.21 ms — under half the 3.33 ms IMU period, so nearest-sample lookup is unambiguous.

TAXEL LAYOUT · DERIVED, NOT VENDOR-SUPPLIED

The pressure-map hand places every taxel at its physical position: 4 rows × 3 columns per digit (tip → base), one knuckle cell per digit, and a 5-row palm pad (12 + 4×15 cells; the three thumb-side cells of the top row are absent). This layout was recovered from the glove recorder's raw 16×16 scan-matrix export (EP-04) by matching each region-array cell to its matrix column over time — not from a vendor drawing. EP-01…03 were captured with the earlier CSV format that omits the raw matrix; the same array ordering is assumed there (same glove model / firmware region extractor), so their in-region positions are inferred rather than measured.

EP-05…23 (stereo rig) ship a 23×20 scan matrix addressed only by electrode crossing (r05c11) with no region mapping supplied. The mapping used here was recovered from the vendor's own quad-view render: each drawn taxel dot's brightness was correlated against all 460 CSV channels, then the resulting positions were clustered — 206 dots per hand matched at r > 0.9, agreeing with the render's six labelled region traces on 97–99% of points. Recovered layout: palm = rows 0–14 × 16 columns (240 cells); each digit = rows 17–22 × 4 columns (24 cells), mirrored between hands; rows 15–16 and the four unused palm columns are padding and read exactly zero in all 38 gloves checked.

ANNOTATION COVERAGE

Two annotation layers, with different coverage. VLM-schema action segmentation — the right-hand rail, the semantic bar and the timeline marks — now covers EP-01 and all nineteen stereo-rig episodes, EP-05…23 (39 segments on EP-01; 5–32 per stereo-rig episode, 406 in total). Canonical-primitive rich labels — the per-arm lanes under the video — cover EP-01 (60), EP-02 (68) and EP-03 (45), on the canonical-14-primitive v1.2 vocabulary plus box-manipulation extensions. EP-04 has neither.
Provenance of the EP-05…23 segments. These are Claude drafts awaiting expert sign-off, not vendor or pipeline output. Method: a visual pass over 0.4–2 fps contact sheets from cam0, with every segment boundary snapped to a hysteresis contact edge on the per-hand taxel-force sum — because this rig is hardware-synced, those boundaries are exact to the frame rather than estimated. Verb, object and target are read from the video; grasp type is the weakest field at this camera angle and carries the lowest per-segment confidence. Segments cover 57–82% of each episode's duration: the uncovered remainder is transit and no-contact time, which is deliberately left unlabelled. EP-01 segment frame numbers are referenced against the source session's raw frame count (3,698) rather than this page's 3,657-frame video; indices are clamped.

EPISODE

GLOVE RIG STEREO RIG

CAM-B | CAM-C · SKELETON OVERLAY

NO CAMERA STREAM IN THIS CAPTURE

Glove telemetry only — dual-hand taxel pressure + IMU orientation, replayed on an internal 30 fps clock. Recorded with the standalone glove recorder (per-hand CSV export).

DEPTH

LEFT IMU—

ACCEL m/s²

GYRO rad/s

ROLL

—

PITCH

—

YAW

—

MANO HANDS21 joints · mm · camera frame

	WRIST	THUMB	INDEX	MIDDLE	RING	PINKY
RIGHT
LEFT

MANO hand pose is available on the glove-rig episodes (EP-01…03) only. The stereo-rig captures (EP-05…23) ship no hand pose and no finger bend sensors, so hand configuration is not recoverable from telemetry alone — only contact location and force are.

RIGHT IMU—

ACCEL m/s²

GYRO rad/s

ROLL

—

PITCH

—

YAW

—

FORCE / TOUCHraw taxel-sum, a.u.

REGION	L	R
Thumb
Index
Middle
Ring
Pinky
Knuckles
Palm

LEFT

RIGHT

LOW

HIGH per-taxel · physical layout · zero-offset per cell

00:00 / 00:00

CURRENT SEMANTIC ANNOTATION

—

FRAMES

—

CONFIDENCE

—

GRASP / CONTACT / MOTION

—

OBJECT / TARGET

—

STATE CHANGE

—

SEGMENT 0/0 · —

click timeline to jump

EPISODE

DURATION

GLOVE PAIRING