Morphospace Atlas
Every tile holds a different evolved soft body, swimming in place. A search called MAP-Elites filled the map by keeping, for each combination of how a body turns and how much it pushes loose particles around, the most robust design it found.
How the atlas works, and what to try
A body is a small soft robot, a grid of up to 8 × 8 squares, each skin, muscle that pulses on a rhythm, or cilia that push along the body’s outer edge. It swims through thick fluid among loose particles. Bodies come from two kinds of genome: a patterned one (a CPPN, a small network that paints tissue by position, so bodies tend to be symmetric and regular) and a direct one that sets each square separately.
The map sorts bodies by two behaviors. Left to right, how much a body turns, from straight runners to fast spinners; bottom to top, how much it moves the particles around it, from leaving them alone to bulldozing them. The search, MAP-Elites, keeps the best body it has found for each square of the map. “Best” means most robust: each body is run several times with its muscle timing, cilia strength and drag jittered, and scored by its worst run, since designs that only work when the biology is precise don’t survive the move to living tissue.
The controls. Switch between the two genomes. Color the tiles by robust activity, or by how well each body does at a task the search never saw: pushing a heavy pellet, holding station in a current, or escaping a crowd. “Replay evolution” fills the map in the order the search found things. Scroll or pinch to zoom, drag to pan, and click a tile to replay that body with its particles and see its scores and the square it came from.
Things to try. Replay evolution and watch the map light up. Zoom into a corner: spinners on the right, bulldozers at the top. Color by “Hold station in a current” and see where the bodies that stay put sit, then compare with the table below the map. Switch to the direct genome and see how much of the map stays empty.
Loading…
Scroll or pinch to zoom, drag to pan, click a tile to inspect its body.
Conclusion
Advocates of open-ended search argue that keeping a diverse collection of solutions finds answers that chasing a single goal misses. We tested that on three tasks neither search saw, comparing the atlas with a search that spent the same effort only on swimming far. So far, it depends on the task. The atlas did much better at holding station in a current: a behavior the distance-chasing search throws away, though the task partly rewards it by design. On the two tasks that reward speed, pushing a pellet and escaping a crowd, the single-goal search matched or beat it. Diversity paid off where the new task needed something the single goal discarded, and not otherwise. The kind of genome made a bigger difference than the kind of search: patterned bodies filled almost the whole map and swam several times faster than bodies set square by square, which left over a third of it empty. These come from a small test search, a 16 × 16 map and two runs of each kind, so they are a first look that a full-size search would need to confirm.
The science
The first xenobots were designed this way: an evolutionary search tried body shapes in a simulator, and the best were built from frog tissue. A design only survived the move to living tissue if it still worked when the biology was sloppy, so here every body is scored by its worst run under jittered muscle timing, cilia strength and drag.
The open question is whether keeping a diverse map pays off. A review of open-endedness in synthetic biology argues that diversity finds solutions a single goal misses. The table above tests that on tasks neither search saw.
What’s real
- Xenobot bodies were designed by evolutionary search in a voxel simulator, then built from frog tissue.
- Designs that tolerated imprecise contraction were the ones that worked in living tissue.
- Ciliated and muscle-driven bodies are both real: later xenobots swim with cilia.
What’s simplified
- Bodies are 2D instead of 3D, and tissue is a grid of springs instead of real cell mechanics.
- Tissue strengths, drag and the fluid model are invented, not measured.
- Particles are simple disks; the surprise tasks are toy versions of real jobs.