Chip Archipelago

Can a biobot be trained for a hard environment by putting it through easier ones first? Here soft bodies and microfluidic chips evolve together: chips mutate into new ones, each with a resident body optimized for it, and bodies move into other chips when they do better there. Then fresh bodies are optimized directly on the final chips, to see whether the stepping stones mattered.

How the archipelago works, and what to look for

A chip is a channel, up to six segments long, that a body swims through from the inlet on the left toward the outlet. In the view, walls and pillars are gray, sticky zones amber and slippery ones blue; a current along the channel and swirling vortices act on the body but aren’t drawn. A body is a small soft robot, an 8 × 8 grid of skin, muscle and cilia. Its score, progress, is how far along the channel it gets in 15 simulated seconds: 0 at the inlet, 1 at the outlet.

POET grows a set of chips and bodies together. It starts with one chip and one body. Each iteration, every body is improved on its own chip by trying random variations and keeping the best. Every 10 iterations:

chips whose body gets at least a third of the way mutate into new ones. A new chip is kept only if it is neither too easy nor impossible for the current bodies (some body gets past 0.2, none past 0.9) and unlike the chips already made;
bodies are tried in each other’s chips and move in where they beat the resident: a transfer;
with more than six chips (twelve in the longer runs), the oldest retire.

At the end, the test: a fresh body is optimized directly on each final chip, with the same number of iterations but no stepping stones. If POET’s residents do better, the path through easier chips mattered.

On the page, pick a run (each seed is one POET run; seeds 24–26 are the longer ones). The live view replays a chip with its resident body, or with the directly optimized body for comparison, or with any other chip’s body swapped in. Below it, the archipelago shows every chip the run made, from creation to retirement, with the transfers between them; then the selected chip’s curriculum, its size as a real chip, and the results across all runs.

Things to look for. The page opens on the final chip where POET did best against direct optimization: run the resident, then the directly optimized body, and compare how far each gets. Follow the blue curriculum back through the chips it came from. Swap in another chip’s body to see how specialized the bodies are. Then try seed 15, where direct optimization won on every final chip, as a reminder that the stepping stones don’t always help.

Loading the POET runs…

The result

No clear advantage at this scale. We ran POET 24 times, 80 iterations each, and compared its residents with bodies optimized directly on the same final chips for the same number of iterations. On the first 12 seeds POET looked clearly better: ahead on 10 of 12, with a mean gain of +0.12 ± 0.04 in progress (standard error). But those were the seeds we used to settle the method. Twelve fresh seeds, run afterwards with nothing changed, came out even: POET ahead on 6 of 12, mean gain −0.02 ± 0.05. Over all 24 seeds the gain is +0.05 ± 0.03, which could easily be chance.

What does hold up: POET builds real curricula, chains of up to 9 chips, and on some chips its residents do far better than direct optimization. On seed 9’s chip 9, POET’s resident got 0.88 of the way and the directly optimized body 0.15. Over all 143 final chips, POET was ahead by more than 0.3 on 25 and behind by that much on 12. On other seeds the stepping stones led nowhere useful: on seed 15, direct optimization was better on all 6 final chips.

Two fixes came first, applied to POET and the control alike. In the first runs each search started from one random body, and on some seeds that body never got far enough to make a new chip, so POET never started. Now every fresh body is the best of 16 random ones on its chip, and a body still stuck after 30 iterations is replaced with a new one. No run stalls now. The restart rule was added after seeing a stalled seed among the first 12, which is one reason the fresh seeds are the fairer test.

Longer runs didn’t change it. Before running anything, we wrote down a test for a longer setting: 300 iterations, 12 chips at a time and 20-second runs, on seeds 24–26. If POET came out ahead on all three with a mean gain of at least 0.10, we would run 12 more. It didn’t pass: the gains were −0.00, +0.06 and +0.02 (mean +0.02). The longer runs built much longer curricula, chains of up to 18 chips, but no final chip showed a gap of more than 0.3 either way. On seed 25 the chips ended up easy enough that both methods solved most of them; on seed 26 both stalled at about the same point.

Preliminary: these are runs on a laptop, with small chips and one setting. The original POET paper found its gap on environments much harder than direct optimization could solve, over thousands of iterations; chips that start out out of reach are the open question.

Conclusion

We asked whether training a biobot through a chain of easier chips, the way POET builds its curricula, prepares it for a hard chip better than training on that chip directly. At this scale, mostly not. POET does build real curricula, chains of up to 9 chips (18 in longer runs), and on some chips its bodies do far better than direct optimization. But averaged over runs, the advantage seen on the first twelve seeds disappeared on twelve fresh ones, and a pre-registered test of longer runs found no gain either. The first seeds had also been used to settle the method, which is why the fresh seeds are the fairer test. One likely reason is that these chips are within reach of direct optimization: the original POET found its advantage on environments that direct optimization couldn’t solve at all, over thousands of iterations. Whether stepping stones help with chips that start out of reach is the open question. The curricula themselves are something concrete: an ordered list of chips, in approximate real sizes, that a lab could put bots through.

The science

POET (Paired Open-Ended Trailblazer, Wang et al. 2019) keeps a population of environment–agent pairs. Each agent is optimized in its own environment; every few iterations, environments mutate, and a child is kept only if it is neither too easy nor too hard for current agents (the minimal criterion) and is novel compared with every environment seen so far. Agents are also tried in other environments and move in when they beat the resident. In the paper, walking robots trained this way solved obstacle courses that direct optimization couldn’t, because the path to the solution ran through other environments.

Here an environment is a chip: a channel up to six segments long with its own widths, pillars, a current along it, vortices and sticky or slippery zones. A body is a soft body from the Morphospace Atlas, an 8 × 8 grid of skin, muscle and cilia grown by a small neural network (a CPPN). Its score is how far along the channel it gets in 15 simulated seconds; the same simulation runs here in WebGPU and in the research code in JAX, and the two are checked against each other.

One change from the paper: its optimizer (an evolution strategy that follows the average of many small random steps) barely moved here, because a body’s tissue map changes in jumps, so its score is flat for small steps. Instead, each iteration a body makes 16 random variants with step sizes drawn between small and large, and the best replaces it if it does better. Each search starts from the best of 16 random bodies, and one still short of the reproduction threshold after 30 iterations starts over from a new one.

What’s real

POET and minimal-criterion coevolution are established open-endedness algorithms.
Microfluidic chips are a standard way to put cells and small organisms through controlled environments.
Biobots built from cilia-covered tissue swim, and their shape changes how they move.

What’s simplified

The soft bodies are a 2D abstraction of tissue, not cells.
Flow is a prescribed field, not solved fluid dynamics; walls are soft.
Chip sizes are approximate (1 simulation unit ≈ 1.5 mm), not a fabrication-ready design.
These are short runs: 80 iterations and six chips at a time (300 and 12 for three longer seeds), where the paper ran thousands.