Cookies help us deliver our services. By using our services, you agree to our use of cookies. Learn more

crowdAI is shutting down - please read our blog post for more information

Learning How to Walk

Reinforcement learning environments with musculoskeletal models


Standing still with Keras-RL

Posted by ViktorF over 4 years ago

I think just standing still would be a good starting point.

The resulting model should be able to balance the body indefinitely, even in the presence of small disturbances, like added noise to the observation vector.

This is what I’m working on right now. Not quite there yet, but I have some promising results already…

Posted by spMohanty  over 4 years ago |  Quote

Well, the idea of “balancing” can also be seen as a sub-policy that your agent should learn to be able to learn the more general policy of moving forward. In most of the cases, I think it anyway figures out this sub-policy to be able to move forward. (Also, balancing while trying to stay a particular place is different than balancing while trying to move, or take a step, or land while taking a step, etc). If you try to custom design all these sub policies, the overall policy will get really complicated (while still not being exhaustive) in no time in my opinion.

Posted by ViktorF  over 4 years ago |  Quote

You seem to be right in all what you wrote.

My plan was to use the standing position as a more stable starting point (than random initialization) for learning how to walk, but it might indeed be unnecessary.