Growing Anthrobots

One cell grows into a stylized anthrobot: a cluster first, then a skin layer and an outer coat of cilia. Every cell runs the same small neural network, trained so the body grows, keeps its shape and heals when cut. Once the cilia are out, they roll the body around the dish.

How the growth works, and what to try

The body grows on a 3D grid, 24 cells across. Every cell runs the same small neural network: each step it looks at its neighbors and updates its own state, sixteen numbers of which four are visible (alive, and whether it is interior, skin or cilia) and twelve are hidden, free for the rule to use as it likes. The rule was trained to grow three shapes from one cell (round, elongated and irregular), keep them, and regrow them after damage. Once cilia cover the surface, they push the body around the dish.

The controls. “Restart from one cell” grows a new body; the knife cuts it with a drag or wounds it with a click, and “Cut in half” removes half at once. Body shape picks one of the three trained shapes, or a blend with the sliders; it changes growth and regrowth, so restart or cut to see it. The view shows tissue or any one hidden channel, and the cross-section looks inside. The rolling settings change how the cilia push: floor grip, which hidden channel sets the beat direction, and whether cilia beat along that channel’s contours or down its gradient.

Things to try. Restart and watch one cell bloom into a cluster, grow its skin and turn its cilia out. Cut it in half and watch it re-seal and regrow. Switch the view to a hidden channel during growth or healing: these are the channels that turn out to foretell what each cell becomes (the research test below). Choose the elongated shape and restart: lopsided bodies roll in circles. On a round body, lower the floor grip to see it spin in place, or raise it to roll it in a straight line.

Loading…

The research test: is there a prepattern?

Do the hidden channels know what a cell will become before the body shows it? We grew 128 bodies under random shape settings, and cut 128 more, then trained simple decoders to predict each cell’s final tissue from its state at each moment: once from its hidden channels, once from its visible ones. Every score is on bodies the decoders never saw.

Loading…

Yes, with a clear lead after a cut. While a body grows, its hidden state predicts each cell’s fate about 8 steps before its visible state does (R² 0.8 at step 20 against step 28). After a cut, the cells that have to change, such as interior cells left on the wound surface that must become skin and cilia, carry their new fate in their hidden state well before they show it (R² 0.68 against 0.44). That lead holds even when the hidden decoder gets only 4 inputs, the same as the visible one. During growth, with 4 inputs, most of the lead goes: the hidden channels hold more information rather than one early signal.

Researchers at Tufts have reported something similar in real embryos: patterns of voltage across a frog embryo’s surface that mark out the face before it forms. Here, nothing asked the rule to plan ahead; it did so because that was a way to grow, hold and heal the body.

Conclusion

A single small rule, run by every cell, can grow a body from one cell, keep its shape and heal it after a cut, and in learning to do that it began to plan ahead. Its hidden channels predict what each cell will become several steps before the cell shows it, and after a cut, the cells that must change carry their new fate well before they change. Nothing in the training asked for this; it was simply a good way to build and repair a body. That echoes the voltage prepatterns seen in frog embryos, and hints that such hidden blueprints may come naturally to anything that has to build and rebuild itself. The same grown bodies also show that, in this model and with the default beating, shape largely decides how a body moves: lopsided ones circle, round ones spin or roll straight. The caveats are real: one trained rule, hand-drawn target shapes, a grid of 24 cells, and hidden channels that stand in for cell signaling without modeling it, so the prepattern is a property of this learned rule, not a claim about anthrobots.

The science

Anthrobots assemble themselves from adult human airway cells. Grown in a gel, the clusters keep their cilia, the beating hairs that normally sweep mucus, on the inside; moved to a thinner medium, they turn inside out, cilia facing outward, and start to swim. Some circle, some go straight, some wander.

The body here grows from a neural cellular automaton: a 24×24×24 grid in which every cell holds 16 numbers. Four are visible (alive, skin, cilia, interior); twelve are hidden, free for the rule to use however it likes. Each step, every cell looks at its neighbors and updates itself with the same small network, trained by gradient descent to grow three target shapes from one cell, hold them and regrow them after damage. Two extra inputs that every cell sees pick the shape; settings in between blend them.

Rolling is a separate, simple physics model. Each cell with cilia pushes along the body’s surface, in a direction set by one hidden channel; the pushes add up to a force and a twist that move and turn the body through thick fluid, with no momentum, as at the scale of cells. In this model the shape decides the path: elongated and lumpy bodies circle, while round ones spin in place or roll in a straight line, depending on how much they grip the floor.

What’s real

Anthrobots self-assemble from adult human airway cells, turn cilia-out and swim; their paths include circles and straight lines.
Neural cellular automata are an established way to learn growth rules that build, keep and regrow a shape.
Bioelectric prepatterns that precede anatomy have been reported in frog embryos.

What’s simplified

The target bodies are hand-drawn (layers by depth), not measured from microscopy; the grid is 24 cells across.
Rolling is a rigid body with sphere-like drag and an assumed beat direction, not real ciliary beating in fluid.
The prepattern test uses one trained rule and per-cell decoders that don't see a cell's neighbors.