Calibrating hand tracking to your own hand — with real data from a live session

Symptoms that call for calibration

#

Real data: what a live session produced

#

MuJoCo’s viewer has a Control panel showing the live force of every motor. Since the force is \(F = 50 - 100 \cdot \text{flexion}\), every recorded frame gives back the exact flexion the tracker published: flexion = (50 − F) / 100.

Previous Next

Reading it like a calibrator:

• The open hand is well calibrated — fingers at 0.02–0.06. The thumb at 0.20 means this operator’s relaxed thumb sits about 0.12 rad below THUMB_STRAIGHT_ANGLE; lowering it by ~0.1 brings the open thumb to ~0.
• The fist never reaches 1.0. Fingers top out at 0.67–0.76, the thumb at 0.61 — the fist angles sit above the curled limits. Raising RAW_CURLED_ANGLE (and THUMB_CURLED_ANGLE) toward the measured fist values would use the full range.
• Middle and index curl less than ring and pinky in the same fist (0.67 vs 0.76) — an argument for per-finger windows.
• The twin still looked right in every pose, because anything above ~0.51 already closes a finger. These gaps only become visible once actuation is proportional.

The procedure

#

1. 1 · Print the raw angles

   Temporary diagnostic in get_finger_flexions

   for finger_name, triplets in self.finger_triplets.items():
       angles = [self._calculate_angle(lm[p1], lm[p2], lm[p3]) for p1, p2, p3 in triplets]
       avg_angle = np.mean(angles)
       print(f"{finger_name:>6}: {avg_angle:.2f}", end="  ")   # TEMPORARY

   Add a bare print() after the loop to end the line.

2. 2 · Run it

   Standalone for the fastest feedback

   env -u PYTHONPATH venv/bin/python standalone/main.py — or in containers, docker compose up vision_tracker and docker compose logs -f vision_tracker.
3. 3 · Measure open

   Relaxed, not hyperextended

   Hold your hand the way you’ll hold it while operating. Read the value each finger settles around — typically 3.0–3.1 for fingers and 2.7–3.0 for the thumb. These become RAW_STRAIGHT_ANGLE and THUMB_STRAIGHT_ANGLE.
4. 4 · Measure closed

   A comfortable full fist, thumb tucked

   Typically 1.5–1.8 for fingers and 2.1–2.4 for the thumb. These become RAW_CURLED_ANGLE and THUMB_CURLED_ANGLE.
5. 5 · Write the constants — in every copy

   Leave a 0.05 rad margin

   Set the straight limit ~0.05 rad below your open reading and the curled limit ~0.05 rad above your fist reading, so relaxed poses reliably clip to exactly 0 and 1. The constants live in standalone/main.py, vision_tracker/src/vision_tracker_node.py, and the guard in docs/tools/make_figures.py. For the container, docker compose restart vision_tracker — no rebuild.
6. 6 · Remove the print

   Printing 30 times a second costs real time in the loop.

Going further: per-finger windows

#

The data above already argues for it:

CALIBRATION = {                     # (straight, curled) in radians — measure your own
    "thumb":  (2.80, 2.50),
    "index":  (3.10, 2.10),
    "middle": (3.10, 2.10),
    "ring":   (3.10, 1.95),
    "pinky":  (3.10, 1.95),
}

(Values back-computed from the live table above — e.g. a finger fist flexion of 0.67 means an average angle of 3.10 − 0.67 × 1.50 ≈ 2.10 rad. Measure your own rather than copying.) In production these belong in a ROS 2 parameter file, so recalibrating never touches source code.