The thing
A polymer pavilion fabricated by robotic arm at RMIT, exhibited at the National Gallery of Victoria. Geometry generated through agent-based behavioural systems; tool-path planning written from scratch in C# and Python; fabrication on a 6-axis ABB robot extruding non-planar polycarbonate.
The problem
Generative architecture has been stuck in renderings for two decades. The geometric language that emerges from agent-based systems — fibrous, layered, self-organising — is unbuildable with any standard fabrication method. Either the architecture domesticates itself to match what a CNC mill can produce, or the discipline keeps publishing pretty pictures.
Floe was an attempt to close the gap on a small, defensible scale: real polymer, real robot, real geometry the design tools described, no manual remodelling between the simulation and the toolpath.
What I did
Three things, in order:
- Behavioural geometry generation. Wrote agent-based scripts in Grasshopper + custom C# to grow the surface field. The output was a non-manifold soup of polylines, not a clean mesh.
- Toolpath translation. Built a custom translator from the agent output to a 6-axis robot toolpath, including non-planar layering, collision avoidance against the build plate, and a print-order optimiser.
- Fabrication & QA. Ran the actual print job. Tuned extrusion rates and motion speeds against polymer behaviour. Documented the failure modes for the next student.
What broke
Polymer cools faster than it bonds at high motion speeds. The first three runs delaminated at the layer interfaces. The fix was a non-planar dwell pattern — the robot pauses on edge nodes long enough for bonding before transitioning to the next path segment.
Why this matters
The Floe work is the spine of how I think about computational design today: a generative system is only as interesting as its proof. If you can't fabricate it, demonstrate it, or ship it — it's a render.
The discipline that goes from agent simulation to robot toolpath without a human in the middle is the same discipline that goes from RAG retrieval to user interface without hallucinating in front of a partner. The translation layer is the work.
Recognition: nominated for Best Master's Student Project (RMIT). Advisor: Prof. Roland Snooks.