A research platform for studying emergent behavior through multi-agent reinforcement learning.

A Spider-Agent learns to maintain a simulated household while an adversarial Saboteur-Agent tries to destroy it - with zero hardcoded heuristics. All decisions emerge from sparse rewards (+1/-1, Baker 2019) and PPO-trained neural networks.

No heuristics, no ontology, no IF-THEN rules. All decisions come from the PPO-trained neural network.