Five populations. One shared objective. Get the cargo across.
Five independent populations evolve cube layouts, arch profiles, cable attachments, girder properties, and car morphology plus a neural driving policy. Their work is then combined into delivery trials. The main viewport is a live, unranked exhibition; training continues independently in a worker.
Each genome gets one solo physics trial and two shuffled-team trials per generation. With the default settings, selection is 0.55 × solo + 0.45 × team. With a population of 12, the best genome is kept unchanged. Tournament selection, crossover and Gaussian mutation generate new candidates, with one random candidate added for variety. Team scores reward two cargo deliveries, progress and speed, with material, energy and damage penalties.
0.55 × solo + 0.45 × team
The highest-scoring sampled team faces two fixed, held-out worlds. The gold curve is the best score on those worlds, not a claim of universal robustness. Archive cells are indexed by girder diagonal stiffness and cable count. A changed world or reward mixture starts a new benchmark; old and new fitness are not compared.
Select Cut cable, then click near a cable. Select Drop load, then click above the structure. Shake disturbs its foundations. These interventions never enter training or validation. R resets the exhibition.
R
Export saves complete populations, random state, best team and the archive at the latest completed generation. Import restores that checkpoint. Optional browser saves use local storage; they may be unavailable in restrictive browser modes.
This is a playable research prototype, not a structural engineering solver. Cubes and lattice cells use particle constraints and approximate contacts. Girders and arch pieces are compliant quadrilateral lattices; cables pull but do not push. Failure uses extension thresholds, not calibrated material stress. Vehicle drive is contact-limited tangential acceleration, not a full tire or torque model.
Builders are evolved parametric construction layouts, assembled with temporary virtual shoring. They are not yet autonomous robot arms, arbitrary graph inventors, or an open-ended technology ecosystem. Their available design families are deliberately specified. One third of the initial genomes are perturbed scaffold priors; the rest broadly sample the gene ranges; no learned weights or prerecorded animation are hidden in the default run.
SPACE: pause replay · N: one generation · R: restart replay · G: strain overlayTraining speed and replay speed are separate. Turbo removes the inter-generation delay; it does not change the physics timestep. A paused training worker may finish its current trial before stopping.