AI Agent Discovers and Controls Self-Organizing Patterns

Summary

Researchers introduce the Artificial Experimentalist, a closed-loop framework that lets an autotelic reinforcement-learning agent choose diverse goals and intervene in complex systems through minimal local perturbations. In CARL, the agent operates on Lenia, a continuous cellular automaton capable of producing life-like self-organizing patterns. CARL discovers stable solitons across varied update rules at a higher rate than heuristic baselines, learns to steer existing solitons, and enables humans to guide them through mazes using high-level directional commands. Training across diverse goals, rules, and initial states supports zero-shot generalization to out-of-distribution conditions. The study suggests a route to AI systems that can autonomously investigate and control emergent phenomena.