Reinforcement learning with musculoskeletal models | Hacker News Reader

Reinforcement learning with musculoskeletal models

osim-rl.stanford.edu

131 pointskidzik26 comments

Like literally 3 hours ago I was searching for software that can simulate skeletons and joints to investigate the exact effects of jiu jitsu joint locks, optimal fulcrum points etc. Could this be used or adapted for that? Anyone here ever used it?

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? :-)

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.

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.

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.

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?

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.

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

Anyone into this check out dual quaternions

http://www.chinedufn.com/dual-quaternion-shader-explained/

They are magic.

I tried the 2017 competition, and DDPG never converges for me. Since the 30+ dimensional state space is so large, I wonder are there some ML techniques that deal with dimensionality reduction or just large dimensions in general?

A lot of these would be good submissions to the Ministry of Silly Walks

https://www.youtube.com/watch?v=iV2ViNJFZC8

Wow, would be nice if this was integrated into the OpenAI environments

if you look at the github readme, it is openai gym compatible.

I may be late to the game but this is the first time I've looked at something and thought "yeah well The Terminator is about to happen"

The skeleton is not exactly good marketing.

I was going to suggest that they hook the AI up to QWOP.

But this task is strictly harder and more general than QWOP .

But not strictly as humourous.

For PR purposes.

I was curious if anyone had tried to train an AI to play QWOP. Of course they have: https://www.youtube.com/watch?v=e27TUmMkOA0 (among others)

I wonder if you could get a better result by including other factors in the reward function, like trying to maintain a slight forward lean.

there is a huge chasm between this working in sim, and irl.

can this be used to help children with developmental delays and disabilities?

Reply on news.ycombinator.com