2016 — Modelling non-planar pelvis and trunk rotation in a planar simulation model

Source: full paper text, not abstract only.

Research theme

Felton’s whole simulation programme rests on a 2D (side-on plane) computer model of a fast bowler. But a real bowling action is not flat: the pelvis and the shoulders rotate out of the side-on plane, so the left and right hips no longer sit on top of each other when you look from the side, and neither do the left and right shoulders. A standard planar model is forced to pretend they do. This paper introduces a fix — two massless segments of variable length and orientation, one joining the two hip joint centres and one joining the two shoulder joint centres — and asks whether that fix lets a flat model reproduce a genuinely three-dimensional action.

Method: a 16-segment angle-driven forward-dynamics model of the front foot contact phase only, built in Autolev, customised to one elite fast bowler (subject-specific inertia from Yeadon’s 95-measurement model). A common set of 34 viscoelastic parameters (wobbling-mass springs and foot–ground springs) was fitted by simulated annealing across three recorded deliveries simultaneously, then tested on a fourth, previously unseen delivery.

What they measured

The model was scored against the real recorded deliveries on four things:

• How hard the front foot hits the ground, moment by moment (horizontal and vertical ground reaction force, as a % of peak vertical force)
• How fast the whole body’s centre of mass is travelling at release (COM velocity, %)
• How far the trunk has rotated/leaned, moment by moment (trunk orientation angle, degrees)
• Ball release speed (%)

These four are combined into one overall RMS score, with the convention that 1° of angle error counts the same as 1% error in the other measures.

Findings

1. Fitting three deliveries at once gave an overall agreement of 5.8%, with individual trial scores of 6.0%, 5.5% and 5.9%.
2. Testing on a fourth, unseen delivery gave 5.3% — i.e. the model generalised across deliveries by the same bowler without re-tuning. This is the more meaningful number.
3. Broken down (match 1 / 2 / 3 / evaluation): force 11.9 / 10.7 / 11.5 / 10.6%; COM velocity 0.1 / 0.1 / 0.1 / 0.2%; trunk orientation 0.9 / 0.9 / 0.8 / 0.8°; ball release speed 1.3 / 2.2 / 2.2 / 0.2%.
4. The kinematics (what the body does) are reproduced far better than the kinetics (the forces). Average kinematic component score 0.8%; average kinetic (force) component score 11.2%.
5. The authors defend the ~11% force error by comparison: Allen et al. (2012), using an angle-driven pin-joint model of triple jumping, achieved a best kinetic score of 15% (adjusted for comparison). Pin-joint models lack compliance, so ground reaction force is inherently the hardest thing to match.
6. Explicit scope limit stated by the authors: the massless-segment trick should be used only for movements where the non-planar rotations have a minor influence on overall performance, and — critically — “planar simulation models cannot be used to investigate the cause and effect relationship of out of plane movements”. You cannot build the model this way and then use it to ask what out-of-plane rotation does to ball speed. That would be circular.

This is a model-validation result, not a correlational or causal claim about bowling technique.

What this means for video and motion analysis

No direct coaching cue. This paper says nothing about what a bowler should do. Its value to anyone building or trusting a motion-analysis system is different:

• A “flat” 2D representation of a bowling action is not automatically wrong — but only if you explicitly carry the out-of-plane geometry as extra information. Felton did not assume the hips and shoulders were flat; he measured the pelvis and shoulder separation from 3D motion capture and fed those time histories into the 2D model as driver signals. A single side-on camera cannot recover that separation. So this paper is not licence for single-camera side-on analysis — if anything it is evidence that the side-on projection alone is missing something the model needed to be given from elsewhere.
• Know which quantity you’re trusting. Position, orientation and velocity quantities were reproduced to ~1% or ~1°. Force was reproduced to ~11%. Any pipeline that estimates forces or loads from kinematics inherits an order-of-magnitude larger error than one that estimates speeds and angles. Treat a video-derived “load” number with roughly ten times the scepticism you apply to a video-derived angle.
• ~0.8° trunk orientation agreement is a useful benchmark. If your own video pipeline claims trunk lean to better than a degree from a phone camera, that is a stronger claim than an 18-camera Vicon-driven simulation makes.
• Validate on held-out trials. The single most credible number in this paper is the 5.3% on the fourth, unseen delivery — not the 5.8% fit. A system tuned and reported on the same footage is reporting nothing.

Caveats and limits

• n = 1. One elite male fast bowler. Four deliveries. All parameters are specific to that bowler; none of the viscoelastic values transfer to anyone else.
• Angle-driven, not torque-driven. The model was told the joint angles from the recorded performance; it did not generate them. This tests whether the mechanics add up, not whether the model can predict a technique change.
• Front foot contact phase only (approximately the 110 ms up to ball release), not the whole action.
• Simulation, not measurement. The “result” is agreement between a model and a lab recording, both of which have their own errors.
• The paper contains an evident typographical slip in the discussion — “Failure to do so may result in a simulation model providing insights into the mechanics of a movement” plainly intends failing to provide insights. The intended meaning is the restrictive one.
• ECB-funded.

Relationship to other Felton work

Direct precursor to the 2017 ISCSB paper and the 2019 Journal of Applied Biomechanics paper, which take the same model and same bowler and run the controlled comparison this paper skipped.

TENSION: This 2016 paper reports only the best model variant (massless segments at both hip and shoulder) and concludes the method “was suitable”. It does not report what the plain planar model would have scored, so a reader could reasonably come away thinking planarity was never much of a problem. The 2017 and 2019 papers show it was: the simple coincident-joint-centre model scores 8.9–9.1% overall and 18% on force, versus 5.7–5.8% and 11%. The later work does not contradict 2016, but it substantially sharpens it — the honest reading of 2016 in hindsight is “the fix works”, not “the assumption was harmless”. Read 2019 before quoting 2016.

The model developed here is the same one later used in Felton, Yeadon & King (2020) Optimising the front foot contact phase of the cricket fast bowling action and in the elbow-hyperextension work — so the ~11% force error and the n=1 subject-specificity propagate into every downstream “optimal technique” claim.