Test morphological computation with Ant Envs
When trying to run the ant env, one can see that it exploits a lot of morphological computation. See the videos of the training in this report: https://api.wandb.ai/links/manfred-eppe/suy9r8ey High MC means that it jumps around a lot, and this causes it to flip over quite often, so that it can never reach the goal. It would be great to train it with cleansac+MC as intrinsic reward, so that it minimizes morphological computation. It could be that it jumps less then, so that it does not flip over and reaches the goal better. Would be nice to see the videos when minimizing MC and comparing it.
To run the experiments, just make a copy of this hyperopt config and use it with cleansac +MC:
AntMaze_Open_Mod-small-cleansac_her-opti.yaml
Also, add the MC value as real-time metric animation so that we understand what is going on.