Excited to share our work on Morphogenesis in Minecraft! We show that neural cellular automata can learn to grow not only complex 3D artifacts with over 3,000 blocks but also functional Minecraft machines that can regenerate when cut in half 🐛🔪=🐛🐛
To our surprise, the approach was able to generate diverse and complex interiors and in the case of JungleTemple, the NCA even generates a functional arrow trap, which uses a working redstone circuit!
You might have wondered what happened to the poor caterpillar we cut in half. It regenerated thanks to the power of NCA!
Does it always work? No. When things become very complex and less regular, like the cathedral with 3584 blocks, there can be some interesting artefacts.
Our paper broadens the scope of tasks suitable for NCAs. Although the work is in a simplified 3D environment, we hope the model’s capability of generating increasingly complex 3D entities brings us one step closer to real-life, self-organizing, and regenerative physical artefacts
This work uses our EvoCraft API. We will release the code for our NCA soon as well.
@enasmel and myself are excited to announce our paper "Meta-Learning through Hebbian Plasticity in Random Networks" arxiv.org/abs/2007.02686
Instead of optimizing the neural network's weights directly, we only search for synapse-specific Hebbian learning rules. Thread 👇
Starting from completely random weights, the discovered Hebbian rules enable an agent to navigate a dynamical 2D-pixel environment; likewise they allow a simulated 3D quadrupedal robot to learn how to walk in around 40 timesteps in the absence of any explicit reward.
The random Hebbian network is also able to adapt to damages in the morphology of the quadrupedal robot, while a fixed-weight network fails to do so.