CARL: Teaching AI to Play God in Petri Dishes

Every scientist who has watched a cellular automaton run knows the feeling: you set the initial conditions, hit play, and become a helpless spectator. Conway's Game of Life unfolds, gliders glide, blocks blink — and you can't reach in and poke it mid-simulation.

For complex systems research, this open-loop paradigm is the norm. You seed a simulation, let it run, observe the outcome. But real experimentalists don't work that way. They intervene. They form hypotheses, perturb the system, and observe how it responds. They close the loop.

A new paper from Cvjetko, Hartl, Levin, Moulin-Frier, and Oudeyer asks: can we build artificial experimentalists that do the same?

Meet CARL

The authors introduce CARL (Cellular Automaton Reinforcement Learner), an autotelic RL agent that operates on Lenia — a continuous cellular automaton known for producing stunning, life-like self-organizing patterns (think: blobs, walkers, swimmers, and things that look like they should be alive under a microscope).

Autotelic means the agent sets its own goals. CARL doesn't wait for a human to say "make a soliton that moves left." It generates diverse goals, learns a goal-conditioned policy, and discovers how to intervene in Lenia through minimal local perturbations. This is closed-loop experimental science executed by a machine.

mindmap
  root((CARL Framework))
    Goal Sampling
      Discover solitons
      Steer movement
      Navigate mazes
    Action Space
      Local perturbations
      Minimal interventions
      Real-time control
    Environment
      Lenia continuous CA
      Life-like patterns
      Multiple update rules
    Policy
      Goal-conditioned RL
      Zero-shot transfer
      Out-of-distribution

Three Capabilities

The paper demonstrates three distinct skills CARL acquires, each building on the last:

1. Discovery. CARL discovers stable solitons across a wide range of Lenia update rules at a higher rate than heuristic baselines. It doesn't just stumble on them — it actively searches for them, treating each discovered pattern as a goal achieved.

2. Control. Once a soliton exists, CARL learns to steer its movement direction with remarkably few interventions. This moves beyond discovery into active manipulation — the agent doesn't just find patterns, it drives them.

3. Human-in-the-loop guidance. The trained agents enable real-time human control: a user specifies high-level directional commands ("go left", "go toward the top-right"), and CARL translates those into low-level perturbations. Humans can guide solitons through maze environments in real time.

The trained policies generalize zero-shot to out-of-distribution conditions — different update rules, random initial states, and goal configurations the agent never saw during training.

Why This Matters

This paper points at something larger than a cool demo with cellular automata. It sketches the shape of an artificial experimentalist — an AI system that doesn't just analyze data or generate predictions, but actively probes, intervenes, and discovers the causal structure of complex systems.

The implications cascade:

• Biology: Imagine an agent that learns to steer morphogenesis in developing embryos by applying local chemical perturbations, discovering which interventions produce which forms.
• Materials science: An autotelic agent that discovers how to grow crystal structures by experimenting with temperature gradients and dopant concentrations.
• Collective behavior: An agent that learns to coordinate a swarm of robots by discovering which local interactions produce desired global patterns.

Lenia is a sandbox — a simplified, computationally tractable model of self-organization. But the method is general. The key insight is that goal-conditioned policies trained across diverse conditions generalize to configurations the agent has never seen. That's the hallmark of a scientist, not a lookup table.

Limitations

The paper is honest about what CARL can't do yet. The framework is instantiated only on Lenia — a specific, continuous CA with known properties. Steering solitons requires pre-training. The agent's goals are currently limited to spatial and kinematic objectives (position, movement direction) rather than abstract properties (complexity, entropy, information flow).

The authors frame this as "a path toward" rather than "a working" artificial experimentalist — and that framing is accurate. CARL is a proof of concept that the closed-loop paradigm works, not a deployable laboratory assistant.

There's also the perennial RL question: how much of the performance comes from the autotelic architecture versus the underlying RL algorithm, goal representation, and environment design? Ablations would help here.

Bottom Line

CARL is one of those papers that feels bigger than its experiments. The immediate results are solid — better soliton discovery, effective steering, real-time human control — but the real contribution is the paradigm: treating complex system exploration as a closed-loop RL problem with self-generated goals.

If you're building AI systems that interact with the physical world — robotics, synthetic biology, materials discovery — this paper should be on your reading list. The artificial experimentalist is coming. It's going to be autotelic, and it's going to be messy in the best way.