AI-powered Fitness Training running sweatscience fitness
Backlinks: AI Coaching | Modeling performance | Training
Last year Alex Hutchinson, in his generally excellent Sweat Science series, discovered the four level framework for data analytics:
- Descriptive: What happened?
- Diagnostic: Why did it happen?
- Predictive: What will happen?
- Prescriptive: How do we make it happen?
This isn't news for people who work in data, and yet he concludes that have yet to see prescriptive analysis. I think he's wrong, the research paper didn't do much research, and I said as much on Twitter:
Good to know. There are a bunch of AI-powered training plans starting to emerge these days (e.g. from Training Peaks, SVEXA, and others). The hard part is evaluating how well they work in the real world!
— Alex Hutchinson (@sweatscience) November 10, 2021
In part because I think it would be fun to write the machine learning algorithm (an idea I've been spinning on for years), I set out to list what algorithms are already out there. Turns out, there are a lot! I compiled all of the ones that I found into a big table and I'll highlight the biggest names below which seem to be basing their approach on data.
DONE The sheets link is broken since I migrated off Google Drive
- State "DONE" from "IN-PROGRESS"
- State "IN-PROGRESS" from "TODO"
Quick task to fix it.
TrainerRoad
The leader in this category seems like TR, with their adaptive training system for cycling, all based on power meter data. Miss a workout, fail to hit your power targets, or mark a session as easier than the algorithm expected and it will adjust your future session difficulty in response. Cycling is a great place to start with clear quantitative data (power) on which to based the training. As a part of this research, I have started using TR and am enjoying it.
See more at TrainerRoad
triDot
I've seen these folks mostly from their advertisements, and can't speak much more to how well their system works or whether it's truly adaptive. They certainly mention AI in their materials.
Garmin + FirstBeat
Backlinks: AI Coaching | Garmin
A stalwart in the endurance tech space, Garmin has partenered with FirstBeat (whom they now own, I recall) to build a recommendation engine using a workout library and some adaptation of the Bannister model (reading between the lines in their linked PR). Disclaimer: I have been using Garmin watches to train for running and cycling for many years, and as such have developed a positive bias here.
TRIQ-App
They have some research on sports science here. Maybe they have based their rules engine on data that shows performance improvements?
AI Endurance
First a cache of their science page (linked above).
With a lot of material on their blog and relatively transparent materials from the researcher behind their system (Markus Rummel), this one deserves a mention with the top. Though they don't appear to big yet, it doesn't appear that they are just using AI as a hype word.
Xert
The folks at Xert have been doing this for a long time in the cycling space. They have plans that adjust based on performance, and really they seem like the first in this space. Link to science behind the training advisor (cache). Like AI Endurance, their marketing does not seem to rival TR or triDot.
Others: Ultra Trail Coaching, SVEXA, 2PEAK, TrainAsONE, Training Peaks, Sparta Science, Humango, Althetica, Perform, Vert.run, and more
With all of the others, it's hard to separate marketing from whether an algorithm is actually optimized on training data to improve performance. It would seem like Training Peaks would have the right data, but it's not clear that their entry in the market here (coaching-focused, perhaps to be expected) is based on data or whether it's a just a rules-based system. SVEXA sounds fancy but doesn't have a consumer product.
Not even making the list are popular endurance training platforms that haven't mentioned dynamic training: Today's Plan, Training Tilt, Final Surge, and many many more.
For a more complete list that includes training data collection apps, see the list here (cache).