Researchers have built an “artificial experimentalist” that discovers and controls self-organizing patterns inside Lenia, a continuous cellular automaton known for lifelike digital structures. Rather than setting initial conditions and waiting for a simulation to finish, the CARL system observes the process and makes small local interventions as it runs.
CARL uses autotelic reinforcement learning, in which an agent samples its own diverse goals and learns a policy for reaching them. The researchers report that it discovered stable, mobile structures called solitons across more Lenia rule sets than heuristic baselines. After training, the agent could also change the direction of existing solitons using only a few interventions.
A human can give high-level directional commands while CARL translates them into low-level changes, making it possible to guide a soliton through maze environments in real time. Policies trained across varied goals, rules and random starting states also worked on some previously unseen conditions without additional training. This is a result in a simulated system, not evidence that the agent can manipulate biological organisms or physical matter. Its value is as a controlled test bed for a broader idea: experimental software may explore a dynamic process more efficiently when it can intervene, observe the consequence and adapt rather than running isolated experiments from scratch.