UNH researchers use AI for breakthroughs in rock climbing

As innovators explore the potential benefits of artificial intelligence in science and medicine, research being done at the University of New Hampshire could give rock climbers a better grip on how easy or difficult their next rocky challenge might be. 

Blaise O’Mara, a graduate student in electrical engineering at the University of New Hampshire, recently published a research paper investigating how machine and deep-learning models of AI can standardize ways to grade the difficulty of rock climbing routes. As a New Hampshire native, he started climbing in high school, continued through college and has climbed outdoor and indoor spots around the state, including NH Climbing and Fitness in Concord.

Currently, some climbing gyms rely on employees while others use ratings from customers, and there is not one standard way to determine the difficulty. The degree of difficulty is graded on the “V-Scale” which ranges from V0 for beginners up to V17 for the most difficult routes only found outdoors.

A recent UNH Today article on this research noted, “Grading rock climbing routes — assigning a numerical degree of difficulty — is notoriously subjective, relying on personal judgment that can lead to inconsistencies and bias. Could artificial intelligence level the field?”

That was the question O’Mara and fellow researcher Dr. MD Shaad Mahmud set out to explore.

“I might go to New Hampshire Climbing and Fitness in Concord and be climbing a v4 there, and when I go to Indoor Ascent in Dover, I might only be able to climb a v2. So there’s those localized differences in bouldering route difficulty, and that must mean there's some bias that is implicitly set,” O’Mara explained.

Rock climbing, bouldering and speed climbing have all grown in the U.S. and across the globe as more gyms have opened for people looking to improve their strength and agility in not-so-competitive settings. However, for competitive purposes routes have to be graded on levels of difficulty, and as climbing takes hold on the international stage the need for consistency is essential to ensure fair competition.

O’Mara said advanced climbers sometimes lose track of the small differences between lower-difficulty routes and may think a route is easier than it is for someone less experienced. There are a lot of factors to consider, such as the size, shape and type of holds, the distance between them, the angle of the wall, and the skill, complexity and flexibility of moves required to get to the end.

Shaad Mahmud, an associate professor of electrical and computer engineering at UNH, started working nearly two years ago with O’Mara using emerging deep-learning model technologies to eliminate bias from rock climbing grading and the results have yielded breakthroughs for the sport. This field was largely unexplored before they started, and this research is unique as it fuses software and hardware for its assessments of difficulty.

“All the models that we have started work on either image processing or somebody manually doing it, the route method, whatever methods we have, all of them have this kind of traditional image-based or natural language processing based method. But we are building sensors that integrate the rock line, like the pinch and the jug, the sensors would be fused on it, and the machine learning will be able to make decisions based on the data that's coming out of it,” he said.

Using technology at UNH’s Remote Sensing Lab allowed O’Mara to provide these algorithms with massive amounts of data points for each hold on routes, distances, and much more. They found that models analyzing the routes, rather than the climber, were most successful at eliminating grading bias.

“That could be a really good avenue for rock climbing gyms in the future to help out their route setters in making routes that are both accessible varied and also consistent in their difficulties,” O’Mara said. “

This technology’s application in real-time could help gyms across the country design multiple walls and routes that cater to different skill levels, shapes and sizes of climbers and could improve the quality of what is offered at climbing facilities. They have not yet reached a level of efficacy and consistency to scale these methods and make them available, but the future applications of the research could make waves in the spor.

“Let's say, you had a phone, it has a camera, you could take a picture of the route that you’re setting, and you could tell the app that, ‘Hey, I'm using these routes in particular in my sequence, what do you think the difficulty is of this route?’” O’Mara said. “It could do some image processing and extract those route-centered features.”

The findings were promising and they indicated that a day might come when these models will standardize climbing difficulties across the board and get rid of the frustration of climbers trying to climb a route only to find it’s much easier or more difficult than anticipated.

Other models, such as the climber-centered approach, were not as efficient at eliminating bias from grading but they did show promise as a training tool for climbers to understand and receive feedback on their climbing habits and skill level.

Eventually, the researchers hope to expand the literature on rock climbing by studying individual climbers more closely to pinpoint which technology can be used to improve rock climbing performance and assess areas of improvement.

Alexander Rapp can be reached at arapp@cmonitor.com