Generative design algorithms

We developed a design algorithm that generates adaptive architecture.

Users start by defining a building site: pick an area on the map, upload site data, or skip the plot and describe the dwelling they want — size, households, height.

A chat assistant asks a few clarifying questions and proposes several room programs: which rooms exist, how big they should be, and how they connect.

Then they pick one variant and run layout. The app shows a 3D preview of rooms fitted into the site envelope. They can remix, edit, compose variants, and export the result as a model — or pass it to the robotic assembly algorithm to deploy on site.

We started with procedural generation: voxel models whose connections are implicit ports on each face, collapsed by a Wave Function Collapse pass — mathematically, rather than by matching tiles by eye.

A small playground let us inspect geometries, assign connections, and export print files until the system could choose every part.

We learned a great deal about assigning and connecting voxel logic. The three-dimensional generations were not satisfying for architecture. That dead end is why the planner now writes a room program, not a tile field.

Once a site is selected, the system builds an envelope to occupy — not a rectangular slab. Each voxel is classified from open data or a scan: floor, inside, leftover. Occupancy is a slightly smaller buildable fill, not a surface shell.

Analysis slices that envelope into storeys. Floor heights sit at three metres; a floor is buildable only if its voxels run as a connected island. Columns that do not run the full height are dropped, so the program is set back from a leftover rather than forced into a uniform plate.