Morphology Search

Now for the fun part — can the AI actually guide the design of robots that look like something specific?

The idea is simple: describe what you want in plain English, like "a kangaroo", "a crab with large claws", or "an ergonomic and stylish lamp", and let evolution do the rest. Robots are generated randomly at first, then the AI looks at each one and scores how well it matches the description. The best ones survive, get slightly modified, and the process repeats — generation after generation — until something interesting emerges.

We tested two ways of managing this process. The basic approach just keeps the highest-scoring robots. The smarter approach, called MAP-Elites, also tries to keep a wide variety of different solutions — not just the best, but the most diverse. Think of it as filling a gallery with many different interpretations of the target, rather than just copies of the same winner.

The results are clear: MAP-Elites consistently produced better and more varied robots across all targets we tested. Even more surprisingly, the AI was able to come up with its own creative categories to organize the robots — for example, classifying lamps by style or crabs by posture — without being told how to do it. All of this was achieved with very modest computing resources, typically evolving between 100 and 300 robots over 5 to 15 generations. The visuals below show some of the results.

The lamp-like morphology has shown to be a quite easy target since the general lamp architecture is simple for our virtual environment. Indeed, our morphology are composed of long cylinder sticks and ellispoid that can recursively develop from a base. This basic morphology approach make some targets almost impossible to reach by nature.

We started with the (mu + lambda) strategy, which is more simple and a good test for verifying if the global idea is working. We gave a computer budget of 10 generation with 20 individuals each and and random init population of 40 individuals. Every generation we kept the top 5 best ever individuals.

We then switched to MAP-Elites, which changes the game entirely. Instead of just keeping the best robots, it organizes them into categories (features descriptors) — and since the judge is an AI, those categories can be surprisingly creative. Rather than sorting by measurable physical properties, we let the AI classify each lamp by its type: bedside lamp, spotlight, decorative lamp, reading lamp, and so on. Each category then runs its own little competition, pushing evolution to find the best version of each lamp type rather than converging on a single winner. The result is a whole collection of diverse, distinct lamp designs — each fitting the target in its own way. We gave the exact same compute ressources than from the (mu + lambda) experiment to compare the approaches. And the final "lamp" were way better.

To verify the previous results, we ran the simulation with other targets, for example looking for tree-like morphologies.

We also tried asking for trees — specifically, "a tree that captures light". The AI managed to evolve a surprisingly diverse collection of tree-like shapes, which is already a win. However, this experiment revealed an interesting limitation: when we asked the AI to sort the trees by biome (cold, temperate, or tropical), it struggled. The problem is that the robots don't look very realistic, and even a human would have a hard time telling a virtual tropical tree from a virtual arctic one based on abstract shapes alone. The AI needs enough visual context to make those kinds of comparisons — and a low-fidelity simulation doesn't always provide it.

During the experiments, we tried multiple configurations and targets that are important to mention to keep track of them.