Darwin-Swarm

Can intelligent collective coordination emerge purely through evolutionary mechanics without gradient backpropagation? Darwin-Swarm is an artificial life and neuroevolutionary simulation sandbox. A population of lightweight feedforward neural policies undergoes selection, crossover, and mutation inside a shared continuous multi-agent environment, measuring whether random chaotic swarms can naturally organize into structured collective movements.

Rather than standard deep reinforcement learning (PPO/MADDPG), each agent's genome is represented by flattened weight vectors of a compact neural controller. In each generation, fixed-lifetime teams are evaluated in PettingZoo's simple_spread environment under seeded conditions. Fitness is awarded based on landmark coverage, inter-agent dispersion, and collision minimization.

• Neuroevolutionary Loop: Elitist selection, uniform crossover, and Gaussian weight mutation applied directly to policy parameter space.
• Multi-Metric Telemetry: Longitudinal tracking of landmark coverage percentage, collision frequencies, spatial variance, and mean Euclidean distance.
• Baseline Validation: Benchmarked rigorously against stationary zero-action baselines and Brownian random walks to verify genuine behavioral adaptation.

Coordination in multi-agent spaces does not require massive parameter counts or gradient calculations. With the right selection pressure, even minimalistic 2-layer policies discover flocking, task distribution, and obstacle avoidance through iterative generational pressure.