Two simulated robots must find the green food patch while avoiding obstacles. Each is controlled by a small recurrent network (GRU with a residual path) that was evolved with a genetic algorithm, and can emit a signal its partner perceives. The orange halo shows a robot signalling; the coloured lines are the paths travelled.
Pick a checkpoint and a seed to watch a recorded episode. The seed sets the starting positions and the obstacle layout.
Paper: arXiv:2610.09280 · Code: github.com/ferdiex/essim2d3d · Weights: xferdie/essim2d3d-controllers