Deep Reinforcement Learning FrameWorks for OpenSim Model

[akhil s anand]

Hello,

I am looking for implementing deep reinforcement learning algorithms on lower limb exoskeleton model developed using OpenSim. Please guide me regarding a feasible deep RL framework which I can integrate with OpenSim for the task.

Regards,
Akhil

[Dimitar Stanev]

This may be of some relevance

https://github.com/stanfordnmbl/osim-rl

[akhil s anand]

hello Dimitar Stanev,

Thank you very much for the reply, I have seen the page and. I am not clear about how keras is communicating with OpenSim. I am a beginner in both Deep RL and OpenSim modelling, Could you enlighten me on this and is it very feasible for any OpenSim models?

Regards,
Akhil

[Łukasz Kidziński]

Hi Akhil,

in the current version of osim-rl you can create a RL environment from any OpenSim model.
Here is one example https://github.com/stanfordnmbl/osim-rl ... les/arm.py but in principle you can make it even simpler -- just setting the model_path in the new class should be enough to have OpenAI gym compatibility of your OpenSim model.
Once it's in that form the OpenSim RL environment will have two functions: reset() and set(). In that form, you can already use OpenAI RL baselines https://github.com/openai/baselines or other implementations, ex. https://github.com/rll/rllab or https://github.com/keras-rl/keras-rl
You will find more information in the README of our repo https://github.com/stanfordnmbl/osim-rl

Best,
Lukasz

[akhil s anand]

hello Lukasz,

Thank you very much for the details. I am bit curios to know, if you are performing any kind of pre processing or additional settings on the OpenSim models to use it with osim-rl.

Regards,
Akhil