# Muscle-up digital twin analysis

## Executive result

Transition height and pull timing are the immediate bottlenecks; the best attempt also uses the largest side-to-side compensation. Attempt 1 ranked best within this video. Its estimated chest center peaked 0.17 m below the detected bar, while the shoulder midpoint remained 0.03 m below it. All three trials ended without a support-over-bar transition.

## Attempt comparison

| Attempt | Window | Peak chest vs bar | Post-grip speed | Min elbow angle | Elbow asymmetry | Peak power proxy | Rank |
|---:|---:|---:|---:|---:|---:|---:|---:|
| 1 | 2.17-5.30 | -0.17 m | 1.71 m/s | 38 deg | 87 deg | 1988 W | 1 |
| 2 | 14.03-16.93 | -0.19 m | 1.44 m/s | 62 deg | 83 deg | 1462 W | 2 |
| 3 | 21.98-24.77 | -0.25 m | 1.33 m/s | 63 deg | 26 deg | 1467 W | 3 |

The ranking is relative to these three attempts only: 50% peak chest height, 20% elbow flexion, 15% symmetry, and 15% upward speed. It is not a probability of success or a population percentile.

## Evidence-linked coaching

### 1. Raise the pull target before adding more transition effort

**Evidence:** The best chest-center estimate remained 17 cm below the bar; the three attempts ranged 17-25 cm below. No support-over-bar phase was observed.

**Practice implication:** Build explosive chest-to-bar height first: high pull-ups toward the lower sternum, band-assisted high pulls, and low-bar transition drills where the chest can travel over the bar without a maximal jump.

### 2. Finish elbow flexion earlier, while upward momentum is still available

**Evidence:** Minimum projected elbow angle occurred 0.18-0.37 s after peak COM height in all three attempts.

**Practice implication:** Cue 'pull, then get the chest over' as the knees/hips finish rising. Use low-bar jump-to-transition drills and slow eccentric transitions to rehearse elbow-to-bar timing before the body starts descending.

### 3. Reduce twist and scissoring without sacrificing height

**Evidence:** Transition-window elbow asymmetry was 87 deg and 83 deg in attempts 1-2; attempt 3 improved to 26 deg but lost about 9 cm of chest height versus attempt 1.

**Practice implication:** Use a symmetrical grip, keep legs closer together through the kip, and film from a true side view. Add controlled scapular pull-ups and matched-side high-pull reps; stop sets when one arm clearly leads.

### 4. Treat the large tuck as a timing tool, not the primary engine

**Evidence:** Mean knee angle compressed to 25-31 deg and ankle spread reached 1.19-1.46 m.

**Practice implication:** Practice hollow-to-arch swings with feet together, then add a compact knee/hip rise. Progress only when the bar path and shoulder rise remain symmetric.

## What each quantity means

- **Direct/video-observed:** source frames, timestamps, detected bar line, 2D landmark locations.
- **Calibrated estimate:** image-plane metres from the stated 1.887 m height and a standing-pose scale; sensitivity is recomputed at ±7%.
- **Model-derived:** MediaPipe 33-point pose/world-z, YOLO 17-point cross-check, Depth Anything relative inverse depth, person segmentation.
- **Proxy:** COM, acceleration, vertical force, energy and power. These are not force-plate, instrumented-bar, EMG, joint-torque, or individual-muscle measurements.

## Uncertainty and validation

- Primary pose: MediaPipe Pose Landmarker Heavy at 60 Hz; all 1543 frames produced a pose.
- Cross-check: YOLO11m-pose at 10 Hz; median cross-model disagreement was 13.6 px on common landmarks.
- Kinematic derivative sensitivity: Savitzky-Golay windows 9/15/21 frames, plus ±7% scale sensitivity.
- Relative depth is independently normalized per keyframe and is useful for scene layering, not metric depth or temporal motion measurement.

## Important limitations

- Single frontal-oblique camera: global depth and out-of-plane rotation are not metrically identifiable.
- Scale comes from stated height and standing pose; perspective/landmark uncertainty is represented by ±7% scale sensitivity.
- MediaPipe world-z is model-derived relative body shape, not calibrated scene depth.
- Depth Anything outputs per-frame relative inverse depth only; values are not metres and are not temporally calibrated.
- Force, impulse, work and power are COM-acceleration proxies, not force-plate or instrumented-bar measurements.
- No individual muscle activation, tendon load, joint torque, or injury diagnosis can be recovered without additional sensors/models.
- Image-plane joint angles are affected by frontal-oblique projection; timing and within-video comparisons are more reliable than absolute anatomical angles.

## Better next capture

Use synchronized side and front views at 60 fps, lock exposure/focus, place the side camera perpendicular to the bar at roughly bar height, include a calibration board or measured rig dimensions, and record strict pull/dip reference tests. Instrumented bar force or force plates would be needed for true force/power validation; EMG would be needed for muscle activation.