A simulation framework for studying how information loss in perception-action loops determines whether an agent can learn to act effectively.
WorldNN models a complete perception-action chain where an organism must push a rock to a target through lossy channels:
Matter (4D state) → Emission → Channel (noise) → VAE → Organism → Action
Every stage is independently controllable: perception quality, channel noise, VAE compression, and organism embedding capacity.
Coordination quality C_i — cosine alignment between the learned policy and the optimal action — predicts task performance with r = -0.87 across 50 conditions. C_i ≥ 0.6 guarantees learning; C_i < 0.5 guarantees failure.
pip install -e ".[dev]"
pytest tests/ -v
python experiments/coordination_quality.py # GPU recommendedsrc/worldnn/ # Core: matter, channels, environment (VAE), organism, training
experiments/ # One script per objective (13 completed)
results/ # Figures and data