Keras implementation of Upside Down Reinforcement Learning

• This is meant to be as small as possible for educational purposes, so it was not meant to be as flexible as the authors' implementation here (in pytorch). At least not yet...
• Behavior function model assumes a gated (multiplicative) function between state and commands in the first few layers. But this should be super easy to change in Keras (have fun playing around).
• Command scaling is hard coded for this environment.
• Run with: python main.py
• Tested with:

conda 4.10.3
tensorflow 2.6.0
gym 0.21.0