Front-Side Mechanics and Sprint Acceleration: When Running Simulation Stops Running

Paper Reviewed

Haralabidis, N., Colyer, S. L., Serrancolí, G., et al. “Modifications to the net knee moments lead to the greatest improvements in accelerative sprinting performance: a predictive simulation study.” Scientific Reports 12, 15908 (2022).
https://doi.org/10.1038/s41598-022-20023-y

---

Sprinting has been studied for decades. Researchers have measured joint angles, ground reaction force, impulses, joint moments, and countless other biomechanical variables. Researchers and coaches have then translated many of those measurements and observations into technique recommendations intended to improve performance.

That sounds like a mature scientific field.

Yet the authors of this study begin with a surprising admission. Despite the volume of research, the amount of data collected, and the abundance of coaching advice, few studies have tested whether deliberately changing those recommended techniques actually improves sprint performance. In fact, they note that virtually no intervention studies have established whether those recommendations produce better sprinting.

Thus this study. The authors set out with an objective to accomplish two very specific goals

1. to explore how hypothetical technique modifications affect accelerative sprinting performance and
2. assess whether the hypothetical modifications support the front-side mechanics coaching framework.

The front-side mechanics coaching framework and its associated cues have become widely promoted in sprint coaching in recent years, commonly accompanied by claims of faster limb recovery, reduced backside motion, and improved acceleration.

This predictive simulation study set out to determine whether a clear and recognizable front-side mechanics technique would emerge when sprint acceleration was optimized computationally. It did not. On top of that, the simulation produced a more consequential and completely overlooked result: the model could not keep running. This raises a more fundamental question that extends beyond sprinting itself.

The Hidden Theory Inside Every Simulation

We often think of predictive simulations as objective because they rely on mathematics rather than human opinion. But mathematics does not eliminate assumptions. It formalizes them.

What’s more, predictive simulations formalize a whole set of human assumptions about movement and locomotion through rigid mathematics.
Every predictive simulation begins with decisions about how the body is represented, how movement happens and is constrained, what defines success, and what equations govern the system. Those decisions determine what the optimization is capable of discovering long before the simulation begins.

Every predictive simulation is already the mathematical expression of a theory of locomotion.

The optimizer does not create movement from nothing. It searches for the best solution within the mathematical world it has been given. The results may be surprising, but they can never be independent of the assumptions used to define that world.

The Missing Question of This Simulation

If every predictive simulation is already the mathematical expression of a theory of locomotion, then a more fundamental question should be asked before interpreting its findings:

What theory of locomotion is this simulation actually expressing?

The paper does provide an answer, although not in those terms.

It begins from the premise that sprint acceleration is governed by the net horizontal impulse generated during stance. According to the authors, improving acceleration therefore requires increasing propulsive impulse, reducing braking impulse, or both. Technique is introduced as the factor that determines how those forces are produced before and during stance.

That is the locomotion model embedded within the simulation.

The optimization is therefore not searching beyond that predefined idea of locomotion. It can only search for the best movement strategy available within a framework that defines sprint acceleration through horizontal impulse production and the techniques believed to influence it.

That distinction matters because everything the optimizer discovers is ultimately constrained by that underlying model.

The Study

This was not another motion analysis study. The authors attempted something much more ambitious.

Instead of observing athletes and interpreting the results afterward, they asked whether a predictive computer simulation could independently discover a faster sprinting strategy during the acceleration phase.

That is an important scientific and practical question because, in principle, it should remove direct coaching opinion from the process. Rather than telling the computer how athletes should run, the researchers allow the optimization process to search for a movement strategy that produces the desired outcome.

Before moving any further, let’s get out of the way what could be perceived as a bigger issue than it really is so we can focus on what actually matters.

As described by the authors, this study used:

a) A three-dimensional musculoskeletal model
b) scaled to an international male sprinter was used in combination with direct collocation optimal control
c) to perform (data-tracking and predictive) simulations of the preliminary steps of accelerative sprinting.

Anyone not directly involved in research would either immediately consider some or all of these factors a major weakness, or simply not even register these important details of how the study was organized. So let’s break them down.

None of them, separately or collectively, present a fundamental issue. Many biomechanical studies begin with a single subject, as is the case here.

Simulations are routinely used to build a case for future research while allowing researchers to investigate questions that would otherwise require larger participant groups, substantially more funding, or experiments that cannot yet be performed directly. The same applies to three-dimensional musculoskeletal models, which have become an established research tool across biomechanics.

The real issue was disguised within the simulation itself, as mentioned earlier, and that is where all of the consequential problems begin.

So let’s start there.

First and foremost, a simulation designed to evaluate someone else’s claims about sprinting has to demonstrate sprinting. That is, sustained human locomotion—a continuous change of support from one foot to the other at sprinting speed.

So what do you do when your carefully constructed sprinting simulation does not continue producing sprinting? Instead of treating that outcome as a reason to reconsider the model’s foundational premises, the researchers added another restriction to keep the motion within the boundaries of running.

Without that restriction, the optimized motion shifted toward movements resembling jumping, hopping, or diving. To prevent this, the researchers required the model to finish the simulated sequence in a position close to the original tracked sprinting motion, leaving it capable of taking another step.

Even with that added restriction, the optimizer improved performance during the first stance phase while performance during the second stance phase became worse than in the original tracked sprint. These outcomes raised new questions about the behavior of the model itself. Instead, the study proceeded to use that same model to evaluate the claims of the front-side mechanics coaching framework.

The Ground Reaction Force Problem

Before examining what the optimizer actually discovered, it is worth looking at the modeling framework from which it emerged.

The predictive optimization grew directly out of the authors’ earlier development and validation of the same modeling framework, reported in a separate paper. Using the same athlete, musculoskeletal model, experimental dataset and previously determined foot-ground contact parameters, the later study performed a new data-tracking simulation that served as the foundation for the predictive optimization.

That model was designed to reproduce the athlete’s recorded sprinting motion by minimizing differences between the experimental data and the simulation, including differences in the kinematics, joint moments, and ground reaction forces. During validation, however, the authors found that tracking the filtered ground reaction force signal altered the simulated movement. Because filtering had artificially extended the force signal at the beginning and end of ground contact, the model positioned the foot closer to the ground and extended the leg earlier and farther n order to reproduce the filtered ground reaction force signal. As a result, the simulated ground-contact phase became longer, with the largest timing differences occurring around touchdown and the end of support.

The authors explained the prolonged ground contact as an artifact of tracking the filtered ground reaction force signal. That explanation accounts for the extended force signal, but it does not address the larger question raised by the result.

The larger problem, however, lies in the assumption behind the tracking objective itself.

Research has consistently shown that faster sprinters produce higher ground reaction forces than slower sprinters. The question is not whether higher forces accompany faster running—that is well established—but whether they should be treated as the cause of faster running or as one of its consequences.

These are fundamentally different assumptions.

If larger ground reaction forces are the result of faster running, then asking a model to reproduce a recorded force profile may unintentionally distort the movement itself. Instead of allowing the modeled actions to produce movement and the corresponding forces, the optimization is directed toward reproducing the recorded force profile first.

Interestingly, once the optimizer was given the freedom to improve sprint performance, one of its earliest adaptations was to reduce time on support. The second stance phase became shorter than in the original data-tracking simulation, moving closer to values reported for elite sprinting.

This result fits more naturally with an understanding of larger ground reaction forces as a consequence of faster movement than with the assumption that reproducing the force profile is what produces faster movement.

Shortening Time on Support

Once the model had been constrained enough to keep it looking like sprinting, the researchers allowed it to change how the hip, knee, and ankle were used. The resulting simulations found different routes to improved performance.

They did not converge on one prescribed technique or produce a clear front-side mechanics pattern. They did, however, reveal a consistent movement pattern. Across the different solutions, ground contact became shorter, the support foot was recovered more rapidly, the leg folded sooner during swing, and the leg did not trail as far behind the body at take-off.

That pattern matters more than whether the simulation reproduced a recognizable coaching model. The optimizer was not instructed to produce any particular movement pattern. These features emerged repeatedly as the model searched for improved acceleration while being allowed to change how the hip, knee, and ankle were used.

The finding therefore shifts the practical question. Rather than asking how to manufacture a particular front-side appearance, it may be more useful to ask how the runner can leave support sooner and prevent the leg from remaining behind the body longer than necessary.

In the Pose Method framework, that is the function of the Pull: to intentionally remove the support foot because natural recoil alone is not enough to continue the change of support effectively.

The Real Question

The most important result of this study is not that the optimization did not produce clear and recognizable front-side mechanics. It is that two different objectives produced two different movement solutions.

When the model attempted to reproduce the recorded ground reaction force signal, time on support increased. When the model attempted to improve sprint acceleration, time on support repeatedly became shorter.

That contrast should not be ignored.

Another point should not be ignored. Before movement can be optimized, it must first be understood. This study set out to analyze movement without first establishing where and how movement begins. That is a far more fundamental question than whether front-side mechanics exists or whether particular coaching cues are effective.

What is actually causing running?

When ground reaction forces are recorded, they are already recorded as the result. When the movement of the legs or arms is described, it is described visually and as already in motion. But where does it begin?

This simulation cannot answer that question. It was never designed to.

The model was not capable of reproducing continuous sprinting on its own. When allowed to optimize freely, it stopped running, attempted to hop, jump or dive to complete its task, and required additional mathematical constraints simply to be able to continue the running movement.

Nor did the optimization begin from first principles.

Before the first optimization was performed, the mathematical model had already adopted a particular explanation of running. The optimization could search only within that explanation. It could not test whether that explanation itself was correct.

That is the deeper limitation revealed by this study.

The simulation was able to optimize movement, but not to determine whether the assumptions defining that movement were themselves correct.

Until that question is addressed, predictive simulations will continue to optimize within existing theories of running rather than optimize running itself.

The optimizer was simply doing its job. Hopping, jumping, and diving were not mistakes. They were mathematically valid solutions to the problem it had been asked to solve.