I've participated in the NIPS "Learning to run" competition last year. It only computes motion in one direction (forward-backward, the model can't go left or right). What turned me off was the fact that object collision was handled poorly. You can't see a model that actually makes effort to avoid obstacles, instead all top solutions go through objects.

That's only one-side of the story. It is actually a very accurate contact model. The problem is that it's also computationally expensive and therefore we reduce the stiffness of objects to make it faster. It's basically a trade-off between the accuracy and speed. In gaming engines you have speed and it looks good, but then it's impossible to generalize for real-life applications (because the contact is inaccurate).

> The problem is that it's also computationally expensive and therefore we reduce the stiffness of objects to make it faster.

Are y'all using penalty methods for the collisions? Which model does it use?

So all 'joints' only move in 2d, am I getting that right?

In the 2017 challenge yes. This year we are using a 3d model.

I know a guy who is interested in that, take a look at this: https://github.com/Eelis/GrappleMap/blob/master/README.md

That's really cool, thanks for mentioning.

Haven't used the software, but how about pitting intelligent agents against each other (incorporating reasonable models for vision, motor control, etc.) and evolving the optimal martial art? :-)

Might find this interesting for extracting info from recorded bouts: https://github.com/CMU-Perceptual-Computing-Lab/openpose

I’ve also been searching for something like this, wondering if it could be applied to horse racing simulation.

Absolutely, however, you will need to construct a muscle model for horses. The human model is a result of decades of research, MRIs etc.