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


Observation and activation time-series

Posted by ViktorF over 3 years ago

Time-series plot of observation and activation values from a 4m walk, then the skeleton fell out of balance at 4.46m:

Horizontal axis is the number of iterations.

5m walk, then falling out of balance at 5.5m (mass center is well ahead of both feet):

The controller returns activation values above 1.0 in my case. Clipping the activation values at 1.0 causes the skeleton to fall almost instantly, so no clipping is done by osim despite the min and max values for all muscle activations are defined as 0.0 and 1.0 in the XML body model.

Posted by Lukasz_  over 3 years ago |  Quote

Very interesting, thanks for sharing! We will keep it as is in the current environment, but this will be fixed in the NIPS competition that we are preparing. Thanks!

Posted by ViktorF  over 3 years ago |  Quote

7m walk