Biomechanics at Scale
Biomechanics at Scale
The Introduction of Pipeline 2.0
Every other week I write an email discussing what I learn launching and growing Reboot Motion. If you would like to receive it directly in your inbox, subscribe below.
When Reboot launched, we thought our reports would change movement analysis.
We thought once people saw the advantage of movement based metrics, they’d instantaneously adopt our philosophies.
We were wrong.
To be clear- nothing has changed our belief that momentum based biomechanics is the way of the future. And plenty of coaches agree with us, and won’t start their day without seeing our reports from the previous night.
But some disagree, and prefer to do their own analysis.
And that’s OK.
In fact, it’s great.
That’s because the very biomechanists and data scientists who are not relying on our reports still use Reboot Motion to drive their organization forward.
Rather than diving into reports, they “get off” the train earlier and use our infrastructure to:
Generate joint angles 1) over time and 2) at key points for further biomechanics research.
Build prediction models with our proprietary energy and momentum based metrics.
Use the data as the backbone of their own reports.
On top of that, teams are turning into power users by:
Uploading years of markerless motion capture data in a single day.
Sending hundreds of videos to analyze from stadiums without mocap installations.
Using us solely as a processing engine…and therefore as leverage for their front office to innovate.
On the business side, this is great. Markerless motion capture in sports is exploding and we are the people best setup to help front offices, data scientists, biomechanists, and coaches turn it into action.
But on the tech side, this demand has a cost.
Scale is Relative
When we pitch MLB clubs, we tell them we can help “do biomechanics at scale”. If they want joint angles, processed metrics, and reports for the previous night’s game in their inbox before they arrive to work, we can deliver.
We thought this was biomechanics at scale. We were naïve.
It turns out, our partner clubs don’t just want data from their game. They want it from every game. On top of that, they want minor league games. They want historical games. They want everything.
It is our job to make that happen.
Enter Pipeline 2.0
To achieve our goal of being the best in the world at movement analysis, we had to make our pipeline more efficient, which meant we had to rebuild it.
And that’s what we did with Pipeline 2.0, which delivers two fundamental changes:
We moved from server-heavy infrastructure to serverless infrastructure, which allows for extreme parallelization.
We switched from using OpenSim for baseline biomechanics functionality to our own model built on top of the open-source MuJoCo physics engine.
The first change is simple to explain. Our partners will see us turnaround more games in less time. Easy.
The second change, however, needs a bit more detail:
Reboot Motion previously built Pipeline 1.0 on top of OpenSim because we viewed it as the gold standard for biomechanical modeling. But over the past few years, we learned there is no such a thing as a gold standard.
Biomechanics in the Wild
OpenSim is an incredible tool in an academic environment. But it is not designed to be used in the wild- where markerless motion capture is the norm.
Markerless mocap technology is incredible. However, it is still 1) different than marker based systems (which OpenSim was designed for) and 2) imperfect. To ensure our partners got the best, most actionable data, we needed to build a more flexible, more resilient model that worked better with the data our partners send us.
And that is exactly what we did.
Pipeline 2.0 uses a Reboot-built skeletal model that is:
More consistent with the degrees of freedom captured by markerless motion capture protocols, and
More resilient to imperfect capture.
Building for Markerless Data
Showcasing how Pipeline 2.0 is built for markerless motion capture is best done with an example:
Lumbar Extension - Flexion
In typical marker-based protocols, markers are placed at key landmark locations on the pelvis to capture anterior - posterior pelvic tilt and lumbar extension - flexion (the degree of freedom between the upper pelvis and lower spine).
Markerless protocols, however, generally only provide key points at the hip joint centers and the shoulder joint centers. There are typically no key points on the pelvis or lower back, which makes it very difficult to capture the same pelvis degrees of freedom mentioned above.
Pipeline 2.0 deprioritizes the lumbar extension - flexion joint angle in favor of a holistic torso extension - flexion joint angle that uses the hip joint centers and shoulder joint centers to measure the tilt of the entire torso (including the pelvis) relative to vertical.
This is not due to us believing hip and shoulder joint centers are more useful in a vacuum. Rather, they are more useful given the input data we receive.
In short, by building our own custom model, we are able to match the dataset our partners provide and deliver better, cleaner, and more consistent metrics.
Building for Resiliency
Not only did we build a model that better aligns with the markerless data we receive, we built one that is more resilient.
Because, as good as today’s technology is, in-game data is not lab data.
There will be inconsistencies from capture to capture, especially from stadium to stadium. We cannot not let good be the enemy of great almost perfect be the enemy of perfect. But we absolutely can adapt.
Below is an example of how we built for resiliency:
Movement-Specific Segment Scaling
Most academic biomechanical modeling approaches keep segment lengths constant for a specific athlete across all motion capture sessions. At the surface, this makes sense: pro athletes do not change shape on a daily basis.
However, even the best markerless mocap systems have variability.
Therefore, while we keep the mass of a player constant across motion capture sessions (based on the player weight data uploaded), we now scale skeleton segment lengths specifically for each movement- not just each player.
By calculating lengths per event within our inverse kinematics algorithm, we match the segment length to the capture, not an ideal lab setting, creating more consistent outputs.
Moving Forward
The primary reason for the rewrite was to find a better way to do biomechanics at scale.
However, once we started the process, we realized MuJoCo also gave us the ability to build our own models, which we could cater to the needs of our partners.
The result is Pipeline 2.0, which offers:
Faster turnaround times,
Cleaner joint angle measurements,
Additional access to underlying biomechanics functionality, and
Increased customizability.
We believe this is a step function increase in our ability to be biomechanics infrastructure. It will make us a better partner to MLB clubs, and it will allow us to expand our toolkit for biomechanists and data scientists everywhere.
Most importantly, it advances our mission of helping others help athletes move better.