This is a project I’ve been contemplating for quite a while! I stumbled across the idea when I was thinking about how hard it is to come up with pleasing and novel boulder problems. First, there are so many degrees of freedom: Wall size, wall angle, hold kinds, hold location, hold orientation … and the many-body problem for holds! And second, we can’t just choose random assignments: Routesetting requires geometric and physiological expertise to consider how the human body can resist gravitational forces and generate momentum from the strangest contact points and angles. And it requires years of sustained visio-motor integration to translate that understanding to actual climbing sequences. On top of that, good routesetting aims for mentally intriguing problems, which requires to estimate the human perception under consideration of the huge variety among climbers! At some point I wondered - If it’s so hard to decompose and abstract this process well enough to build a theory of good routesetting - can artificial intelligence help us? Surely a machine can learn to create new boulder problems, based on boulder problems which have been crafted by the intelligence of the route setter? Can we find more interesting and challenging boulder problems, if we have ratings? Can we go a step further and learn the route specifications just from images of boulder problems? And while we’re at it - If we know how to condense the image of a boulder problem to an interpretable representation of the boulder, can we use images from world cup boulders to understand how top athletes crack the hardest problems in the world? This project is potentially huge!
This is what I have in mind. Let me know what you think!
Phase 1: Generating new boulder problems
- [easy] Using boulders described by interpretable features: Take an existing database (e.g. from the Moonboard App), where boulder routes are given as a subset of holds placed in specific locations of a constant grid. The set of available holds is fixed, as well as the rotation of the holds. Hence we can train a generative model without supervision, or use available user ratings and climbing grades for supervision. We can then synthesise new routes and directly contribute them!
- [hard] Using only images of (complex) boulders. First generate an image collection of boulder problems in a similar setting (e.g. IFSC world cups). Assume we have an infinite set of possible holds and no additional information on the selected holds, their rotation or actual spacing. Now we can play around with unsupervised representation learning, visit a model zoo and try to capture the route from the image. It might be helpful (but less fun) to label some images (e.g. outline holds within the images) to allow for domain-specific segmentation with weak supervision. Alternatively, the entire problem could be simplified by introducing hold classes (“crimp”, “pincher”, “foothold”, “sloper”, “volume”), although this will demand significant labelling work and will most likely fail to capture modern competition-style boulders, where the route setting and holds are extremely diverse.
Phase 2: Extension and Analysis
- Learn models for sport climbing.
- Use discriminative models to distinguish route setting styles
Phase 3: Build powerful tools
- [time consuming] Build an app for route setters and think of ways to (a) customise to specific hold sets and route styles and (b) allow interactive input. Often a route setter will begin with the bottom part of a route and work her way up, testing the moves on-the-fly. It would be nice to have predictions for the most likely next hold, i.e. a time series framework.
- [most exciting, super hard & open ended] Analyse the athlete’s approaches and solutions to the boulder problems! Predict solutions to unclimbed boulder problems or predict the likelihood for success. Or even learn what distinguishes athletes in terms of technique and athletic performance! There are incredibly many cool things to discover here!