Growing 3D Artefacts and Functional Machines with Neural Cellular Automata

Video of more results: https://www.youtube.com/watch?v=-EzztzKoPeo

---

Requirements

• python3.8
• This package automatically install test-evocraft-py (https://github.com/shyamsn97/test-evocraft-py), but for further functionality follow installation steps here: https://github.com/real-itu/Evocraft-py

Installation

For general installation

python setup.py install

For ray tune + mlflow

python -m pip install -r ray-requirements.txt
python setup.py install

Usage

Make sure an evocraft-py server is running, either with test-evocraft-py --interactive or by following the steps in https://github.com/real-itu/Evocraft-py.

Configs

Each nca is trained on a specific structure w/ hyperparams and configurations defined in yaml config, which we use with hydra to create the NCA trainer class.

Example Config for generating a "PlainBlacksmith" Minecraft Structure:

trainer:
    name: PlainBlacksmith
    min_steps: 48
    max_steps: 64
    visualize_output: true
    device_id: 0
    use_cuda: true
    num_hidden_channels: 10
    epochs: 20000
    batch_size: 5
    model_config:
        normal_std: 0.1
        update_net_channel_dims: [32, 32]
    optimizer_config:
        lr: 0.002
    dataset_config:
        nbt_path: artefact_nca/data/structs_dataset/nbts/village/plain_village_blacksmith.nbt

defaults:
  - voxel

Generation and Training

See generation notebook for ways to load in a pretrained nca and generate a structure in minecraft

See training notebook for ways to train an nca

CLI training

python artefact_nca/train.py config={path to yaml config} trainer.dataset_config.nbt_path={absolute path to nbt file to use}

Example:

python artefact_nca/train.py config=pretrained_models/PlainBlacksmith/plain_blacksmith.yaml trainer.dataset_config.nbt_path=/home/shyam/Code/3d-artefacts-nca/artefact_nca/data/structs_dataset/nbts/village/plain_village_blacksmith.nbt

Spawning in minecraft

See generation notebook for more details

Example spawning the oak tree

1. Load in a trainer

from artefact_nca.trainer.voxel_ca_trainer import VoxelCATrainer

nbt_path = {path to repo}/artefact_nca/data/structs_dataset/nbts/village/Extra_dark_oak.nbt
ct = VoxelCATrainer.from_config(
                    "{path to repo}/pretrained_models/Extra_dark_oak/extra_dark_oak.yaml",
                    config={
                        "pretrained_path":"{path to repo}/pretrained_models/Extra_dark_oak/Extra_dark_oak.pt",
                        "dataset_config":{"nbt_path":nbt_path},
                        "use_cuda":False
                    }
                )

2. Create MinecraftClient to view the growth of the structure in Minecraft at position (-10, 10, 10) (x, y, z)

from artefact_nca.utils.minecraft import MinecraftClient
m = MinecraftClient(ct, (-10, 10, 10))

3. Spawn 100 iterations and display progress every 5 time steps

m.spawn(100)

Output should look like this:

Structures

see data directory. To view structures and spawn in minecraft see generation notebook. An example of spawning and viewing the Tree:

import matplotlib.pyplot as plt
from artefact_nca.utils.minecraft import MinecraftClient

base_nbt_path = {path to nbts}
nbt_path = "{}/village/Extra_dark_oak.nbt".format(base_nbt_path)

 # spawn at coords (50, 10, 10)
blocks, unique_vals, target, color_dict, unique_val_dict = MinecraftClient.load_entity("Extra_dark_oak", nbt_path=nbt_path, load_coord=(50,10,10))

color_arr = convert_to_color(target, color_dict)

fig = plt.figure()
ax = fig.gca(projection='3d')
ax.voxels(color_arr, facecolors=color_arr, edgecolor='k')

plt.show()

This should spawn and display:

Authors

Shyam Sudhakaran [email protected], https://github.com/shyamsn97

Djordje Grbic [email protected], https://github.com/djole

Siyan Li [email protected], https://github.com/sli613

Adam Katona [email protected], https://github.com/adam-katona

Elias Najarro https://github.com/enajx

Claire Glanois https://github.com/claireaoi

Sebastian Risi [email protected], https://github.com/sebastianrisi

Citation

If you use the code for academic or commecial use, please cite the associated paper:

@inproceedings{Sudhakaran2021,
   title = {Growing 3D Artefacts and Functional Machines with Neural Cellular Automata}, 
   author = {Shyam Sudhakaran and Djordje Grbic and Siyan Li and Adam Katona and Elias Najarro and Claire Glanois and Sebastian Risi},
   booktitle = {2021 Conference on Artificial Life},
   year = {2021},
   url = {https://arxiv.org/abs/2103.08737}
}