[lambdaone]

I'm sceptical about the usefulness of this for a number of reasons, including the points raised by others about hold recognition etc.

It's also difficult to see how the system would work out the 3D arrangement of holds. It can clearly try to infer 3D relationships from video of an existing climb, but it would be hard for it to work out the 3D position or orientation of any holds not used by the climber in the video.

These two put together make solving the climbing problem even more difficult, because appropriate body positions and moves are often very sensitive to even tiny differences in shape, relative position, and orientation of holds; and occluding volumes, arêtes, cracks and so can make the problem even harder.

But as a simplified demo, this is cool, and I salute it.