Hey ! Based on my observations hidden layers of the neuronal networks use a sigmoid function of activation. In motor application like this simulation of creatures, this function quickly converge towards "lazy movements" which gives contractile creatures or those that hop or swing. It's quite efficient when we want to jump or fly but get complicated when we want to run or climb like a bipede/quadrupede. Dunno if it's complicated to implement but adding functions like ReLu or Tanh can be a great upgrade for the game ! I can't realize you've done this game 7 years ago u did such a great job !