AI That 'Grows Like a Body' Isn't Automatically Better

When engineers build AI that controls a robot, they usually write the design directly — every part spelled out. But some researchers prefer a different approach inspired by biology: instead of describing the robot itself, you describe a set of 'growth rules,' and the robot's body and brain develop over time, like an embryo turning into a body. The hope has always been that this growth process makes AI more adaptable, more creative, and tougher when things go wrong.

This study put that hope to the test. The researcher ran 30 matched experiments per method — same robot, same starting conditions, same amount of computer time — and compared three versions: direct design, a simple generative design, and one that develops over time. The growing version did produce much more variety in the final designs, which sounds great. But when it came to actually performing the task, it didn't reliably beat the direct approach. Not during training, and not in new situations it hadn't seen before.

Worse, the growing version had some real downsides. When researchers made small random changes — the equivalent of a typo in the instructions — the growing design was more likely to break completely instead of improving slightly. Combining two good designs to make a better one (a technique borrowed from breeding) also produced weaker, often non-working offspring. And damage applied early in development was softened, which is a genuine plus, but that same flexibility made it harder to preserve the good traits a design had already earned.

The takeaway isn't that growth-based AI is bad — it's that 'more lifelike' doesn't automatically mean 'better performing.' The researcher argues we should judge these methods on several qualities at once: creativity, stability, toughness, and how well good traits pass down. It's a useful reminder across AI: a fancier approach that looks more natural isn't a win unless it actually does the job better.