GAD Editorial · Published · · English study · Chinese summary · Study dated 2026-10-02
A generated residential floor plate becomes useful when a team can inspect its boundaries, understand its labels and decide what to change. FLOORA, an Autodesk Research preprint submitted on 28 September 2026, addresses that structured-output problem. It generates conceptual multifamily layouts from building massings using a compact domain-specific language. The research repository provides inference code and links to models and data. Original paper · Research repository
The output represents living units, corridors and cores. A living-unit polygon should therefore be read at floor-plate scale; it does not establish the internal rooms of a completed apartment. The official model card excludes code compliance and construction-ready documentation from intended use. Its account of North American design conventions also matters when considering a project elsewhere. Official model card
Define the experiment around the decision
GAD proposes a small acceptance study before a practice assigns the model a production role. Choose an actual question, such as whether a particular footprint merits further residential planning. Prepare a permitted, simplified boundary and a manually reviewed reference option. State which programme assumptions are fixed and which the exercise may explore. Keep irregular footprints in a separate challenge set so they cannot disappear inside a favourable average.
Record the input geometry, units, orientation and generation settings. Save every candidate considered in the comparison, including outputs that fail to parse. Decide beforehand how many candidates may be generated for each input. Comparing one manual option against an unlimited search of AI outputs would make the result difficult to interpret.
Separate machine checks from design review
The first review should ask whether the output can be read consistently. Check that polygons close, labels survive conversion, units remain correct and no spaces unintentionally overlap or leave the boundary. Document any repair, including whether it was automatic or manual.
A second review should examine the proposed organisation. Trace access from each living unit through circulation to the intended core. Identify thin residual spaces, awkward junctions and places where a plausible diagram still leaves the next designer with substantial work. Relevant specialists should determine the project-specific checks needed before any option advances.
Keep those findings separate from visual preference. A neat-looking arrangement can be expensive to repair; an unusual one may reveal a useful alternative. Ask reviewers to explain the design consequence of each issue rather than assigning a single impressionistic score.
Count the work after generation
Measure preparation, review and correction effort as well as generation time. The practical comparison is the effort required to reach a usable next-stage study. Record the most common failure types and the footprint conditions associated with them.
The paper identifies highly irregular massings as a limitation and does not release its training code or human-feedback data. Those boundaries should remain visible in any evaluation report. GAD has reviewed the documentation, not run the model. A sensible adoption decision would name the narrow task it helps, the checks it passes and the conditions that still require a conventional study. Limitations and reproducibility