Virtual Creatures with Neural Nets Evolve Animal-Like Nervous Systems
Developer neuroecology built a "virtual animal" ecosystem simulation in their spare time and presented the results in several steps: each virtual worm is controlled by its own neural network and must obtain enough protein, carbohydrates, and lipids from different foods, with nutritional balance directly affecting survival; organs send neuromodulators to the neural network to guide decisions; the worms must also evade predators—themselves driven by neural networks—and interact with conspecifics. The author initially trained only a few behaviors such as foraging and egg-laying.
Confirmed
- The system is trained with deep reinforcement learning using the PufferLib framework developed by jsuarez; neuroecology and jsuarez had discussed "modeling evolutionary and social behavior with deepRL," believing such methods can tackle questions other modeling approaches struggle to reach.
- After training, the agents' neural dynamics exhibited "animal-like" nervous system structure: individual command neurons dedicated to steering emerged, and sensory neurons spontaneously differentiated into distinct functional types.
- Predators are also neural-network-driven, creating a multi-agent game-theoretic simulated environment.
Why It Matters
- The author notes that even under very simple evolutionary/training setups, neural networks under environmental pressure spontaneously organize into structures resembling real animal nervous systems, offering a low-cost, reproducible experimental platform for studying neural evolution and behavioral emergence.
- The work demonstrates the potential of deepRL + PufferLib for modeling biological behavior, touching on problems hard for traditional methods—such as foraging strategies under nutritional-metabolic constraints and predator-prey co-evolution.
2026-11-21 ~ 2026-11-21 · 7 related posts
Primary sources
- [source] Developer builds virtual animals with neural networks foraging, hiding and surviving — neuroecology · 2026-11-21
- Hobbyist's virtual worms with neural networks reproduce classic optimal foraging research — neuroecology · 2026-11-21
- [source] Virtual worms spontaneously evolve command neurons and segregated sensory neuron types — neuroecology · 2026-11-21
- [source] DeepRL agents trained with PufferLib evolve animal-like neural dynamics — neuroecology · 2026-11-21
- DeepRL + PufferLib could unlock modeling of evolution and social behavior, researchers argue — neuroecology · 2026-11-21
- Can we decode deepRL-trained virtual animals with interpretability tools? — neuroecology · 2026-11-21
- Tiny neural tweaks in simulated animals unspool whole different lifetimes — neuroecology · 2026-11-21