Is This App the Future of AI Workouts?
The newly revamped TrainerRoad cycling app uses an advanced algorithm to create training programs custom-tailored to the user. Our writer gave it a go.
Outside's long reads email newsletter features our strongest writing, most ambitious reporting, and award-winning storytelling about the outdoors. Sign up today.
To the untrained observer, it doesn’t look like much: I am a skinny 31-year-old male in my apartment bedroom, sweating profusely in spandex bib shorts atop half a bicycle. I’ve swapped the bike’s rear wheel for a smart trainer that tracks my cadence, power output, and speed. It’s classic COVID-era indoor exercise in the same vein as a Peloton bike or Zwift. But instead of a live feed of a cycling class or a video game racecourse, I’m staring at a series of blue lumps graphed on my desktop computer screen. The blue lumps represent the target power measured in watts. As a lump grows, I have to work harder. When the lump shrinks, I get a rest. A thin yellow line shows my actual power output as I attempt to complete each interval. An on-screen timer shows me how long until the intensity changes again. Occasionally, white text pops up with some sage advice from a disembodied coach: “Quick legs, high power.” “Find your sit bones.” It’s majorly nerdy, hardcore cycling training being foisted on one of Earth’s most mediocre athletes who has absolutely no race aspirations.
But behind this facade, a sophisticated artificial intelligence–powered training program is adapting to my every pedal stroke. The app I’m using is called TrainerRoad, and in February, the company launched a suite of new features on a closed beta app that it believes can revolutionize how cyclists train. The new technology is powered by machine learning: the idea that computers can be trained to hunt through massive troves of data and suss out esoteric patterns that are invisible to the human brain. The new TrainerRoad algorithm is watching me ride, evaluating my performance and progress, and comparing me to everyone else on the platform. (How many people, exactly? The company won’t say.) This data is then used to prescribe future workouts—ranging from slow and steady endurance work to high-intensity sprint intervals—that are tailored just for me. “Our vision is that in ten to 20 years everyone will have their workouts picked by an AI,” says Nate Pearson, CEO of TrainerRoad.
The idea of using an algorithm to optimize training isn’t exactly new. Louis Passfield, an adjunct professor in kinesiology at the University of Calgary, has been dreaming of calculating his way to a yellow jersey since he was an undergraduate at the University of Brighton around 25 years ago. “I thought that by studying physiology, I could calculate this perfect training program and then, in turn, win the Tour de France,” Passfield says. “This was back in 1987, before the concept of what they call ‘big data’ was even born.”
What is new is the proliferation of smart trainers. In the late 1980s, power meters were inordinately expensive and confined to Tour de France teams and sports science laboratories. Now, more than 1 million people have registered for Zwift, an app where they can obsess daily over their watts per kilo, heart rate, and cadence. Finding a Wahoo Kickr bike trainer during the pandemic has been about as easy as finding toilet paper or hand sanitizer last spring. All these cyclists equipped with laboratory-grade trainers are generating troves of high-quality data that makes researchers like Passfield swoon. “I’m infinitely curious,” he says. “I love what TrainerRoad is trying to do and how they’re going about it. It’s an area I’m itching to get involved with.”
TrainerRoad was founded in 2010 by Pearson and Reid Weber, who now works as CTO at Wahoo’s Sufferfest Training platform. It began as a way for Pearson to replicate the experience of spin classes at home and has evolved into a cutting-edge training app, especially since the smart trainer boom.
What TrainerRoad has done better than competitors is to standardize its data collection in a way that makes it scientifically powerful. There are many more rides recorded on Strava than on TrainerRoad, but they don’t contain enough information to make them useful: We can see that Rider A rode halfway up a hill at 300 watts, but is that an all-out effort for her or an easy spin? Did she stop because she was exhausted or because there was a red light? More than maybe any other smart trainer software, TrainerRoad has built a data collection tool that can begin to answer these questions. There’s no racing. There’s no dance music (thank god). There are no KOMs (regrettably). There’s nothing to do on the platform except workouts. It’s also not for everyone: You log in and ride to a prescribed power for a prescribed time. It is often brutal. You either succeed or you fail. But it’s the simplicity of the format that has allowed TrainerRoad to be the first cycling trainer software to offer this sort of workout.
This pass/fail duality also underlies TrainerRoad’s nascent foray into machine learning. The technology behind the new adaptive training program is essentially an AI classifier that analyzes a completed workout and marks it as fail, pass, or “super pass” based on the athlete’s performance. “At first, we actually tried to just do simple ‘target power versus actual power’ for intervals, but we weren’t successful,” Pearson says. “Small variations in trainers, power meters, and how long the intervals were made it inaccurate.” Instead, TrainerRoad asked athletes to classify their workouts manually until the company had a data set big enough to train the AI.
Humans are quite adept at making this type of categorization in certain situations. Like looking for pictures of a stop sign to complete a CAPTCHA, it’s not hard to look at a prescribed power curve versus your actual power curve and tell if it’s a pass or fail. We can easily discount obvious anomalies like dropouts, pauses, or weird spikes in power that trip up the AI but don’t actually indicate that someone is struggling. When we see the power curve consistently lagging or trailing off, that’s a clear sign that we’re failing. Now, with more than 10,000 workouts to learn from, Pearson says the AI is outperforming humans in deciding pass versus fail.
“Some cases were obvious, but as we got our accuracy up, we found the human athletes weren’t classifying all workouts the same,” he explains. In borderline cases, sometimes a minority of athletes would rate a workout as a pass while the majority and the AI would rate it as a struggle. When presented with the AI’s verdict, the riders in the minority would usually change their opinion.
Armed with an algorithm that can tell how you’re doing on workouts, the next step—and probably the one users will find most exciting—was to break down a rider’s performance into more granular categories, like endurance, tempo, sweet spot, threshold, VO2 max, and anaerobic. These power zones are common training tools, but in case you need a refresher, functional threshold power (FTP) represents the maximum number of watts a rider can sustain for an hour. Then, the zones are as follows:
- Active recovery: <55 percent FTP
- Endurance: 55 percent to 75 percent FTP
- Tempo: 76 percent to 87 percent FTP
- Sweet spot: 88 percent to 94 percent FTP
- Threshold: 95 percent to 105 percent FTP
- VO2 max: 106 percent to 120 percent FTP
- Anaerobic capacity: >120 percent FTP
As you complete workouts across these zones, your overall score in a progression chart improves in the corresponding areas. Spend an hour doing sweet spot intervals—five-to-eight-minute efforts at 88 percent to 94 percent of FTP, for instance—and your sweet spot number might increase by a point or two on the ten-point scale. Critically, your scores for endurance, tempo, and threshold are also likely to move up a bit. Exactly how much a given workout raises or lowers your scores in each category is a function of how hard that workout is, how much training you’ve already done in that zone, and some additional machine learning running in the background that analyzes how other riders have responded and how their fitness has changed as a result.
Here’s what my progression chart looked like after I had used the new adaptive training program for a few days. The plan I’m on now is focused on base training, so, according to the software, I’m leveling up in those lower endurance zones. If I were training for a crit, I’d probably be doing a lot more work in the VO2 max and anaerobic zones—which is why I’ll never race crits.
In the future, TrainerRoad plans to expand the role of machine learning and build more features into the app, including one designed to help athletes who menstruate understand how their cycle affects their training and another to help you forecast how a certain plan will improve your fitness over time. The company is investigating how much age and gender affect the rest an athlete needs and is even planning to use the system to compare different training methodologies. For instance, one common criticism of some TrainerRoad plans is that they spend too much time in the challenging sweet spot and threshold zones, which could lead to burnout. Meanwhile, there’s a large body of science that suggests a polarized approach—a training plan that spends at least 80 percent of training time in Zone 1 and the other 20 percent in Zone 5 or higher—yields better results and less overall fatigue, especially in elite athletes who have lots of time to train. This debate has been ongoing in sports science for years, with no real end in sight. Now that TrainerRoad has added polarized plans, the company may be able to do some A/B testing to see which plan ultimately leads to greater fitness gains. Tantalizingly, we might even learn which types of athletes respond better to which types of training. “The studies that exist are pretty small sample size,” says Jonathan Lee, communications director at TrainerRoad. “We have thousands upon thousands of people.”
The potential for experimentation is impressive, but one of the limitations of machine learning is that it can’t explain why improvements are happening. The inner workings of the algorithm are opaque. The patterns that the AI finds in the training data are so multifaceted and abstract that they cannot be disentangled. This is where the system’s power comes from, but it’s also an obvious restriction. “PhDs usually want to figure out what are the mechanisms that make somebody faster, but we don’t necessarily know,” Pearson says. “What we care about is just the outcome performance.”
But does this actually work? Does adaptive training make people faster than traditional static training programs, like something you’d find on TrainingPeaks, Sufferfest, or even the old version of TrainerRoad? For now, Pearson says it’s too soon to tell. The closed beta program began on February 25 of this year, with only around 50 users, and has been expanding slowly, with new riders being added every week. That isn’t a large enough sample size to detect statistically significant differences yet. “It sounds like a great idea,” Passfield says. “What it needs is to be objectively evaluated against a standard program and, ideally, against a random program. From a scientific point of view, that’s kind of the ultimate baseline: we give you these sessions in a random order, we give you these sessions in a structured order, and then we give them to you in our AI-informed order.”
Here’s what I can tell you, though. The adaptive training is definitely more likely to make me stick with a plan. Back in the fall, I spent a few weeks using TrainerRoad vanilla for the sake of comparison. I found it excruciatingly difficult, because I am not a highly motivated rider. I’m not training for a race or trying to get KOMs on local climbs. Without motivation, the intervals become pointless torture. With the static training plan, quitting put you behind. The next workout was going to feel even harder since you missed part of the previous one. If you fell behind the curve, you had almost no shot at digging out. Now, if I fail a workout, it’s fine. The next one gets a bit easier. When you open up the dashboard, you’ll see a message like this:
In the old version, I had to show up well-rested, focused, fueled, and perfectly hydrated to complete workouts. But this does not always gel with my lifestyle, man. Before COVID-19, I had friends who liked to drink beer and stay up late. I play hockey twice a week. I surf whenever there are waves. I eat fast food frequently. With the adaptive training, all of this is fine. I can drink three beers after hockey and show up for my workout the next day with nothing but McDonald’s in my body. The AI adjusts for the fact that I’m a deeply flawed, suboptimal human, and honestly, it feels so good to be seen.