New variant of my kind-of #alife task. The bots can push/pull treasure and walls. They are controlled by a neural network, trained with sep-CMA-ES to push treasure together.
GIF
The hope was that they would somehow learn to communicate via environment where the stash is, but instead they just push treasure forth and back until they hit it.
Still kinda fun to watch how they learned this forward-backward strategy, and how they turn.
They can see their six neighbours, plus one hex further straight ahead, and have a (very limited) memory.
The score is the number of treasure-neighbours, squared, for each treasure hex, at step 3000.
Web-Player: https://log2.ch/progenitor/build14c/
log2.chProgenitor