How can generative representations be interpreted into physical action, when they were never designed to be executed?
2024 · Group Research Project · SCI-Arc
Advisor: M. Casey Rehm
Collaborators: Termrak, Wenhao Huang, Wei Ding
My role: Robotic path algorithm design, contour extraction
Generative and AI-assisted design tools are great at producing images quickly. Though they generate rich visual structures, density variations, and spatial patterns, they are not designed to act. Fabrication is the opposite. Physical processes require clarity. Every stroke must be continuous, intentional, and irreversible. In the process of production, there is no noise to “smooth out” once it begins. This project proposes a pipeline that carries generative speculation into physical fabrication, and explores what must be abstracted from generative outputs in order for them to become executable. Through algorithmic contour extraction and fill-pattern design, We translate generative representations into layered toolpaths that a robotic arm can perform.
While AI-assisted generative systems enable rapid visual exploration, their outputs are not designed to act. Fabrication demands the opposite: continuity, commitment, and irreversibility. Once production begins, ambiguity must be resolved into instruction. Our intention was to investigate how generative processes might be reinterpreted when they encounter the constraints of material, motion, and time. Within this framework, I worked on algorithmic translation and developed contour extraction and fill-pattern algorithms that reinterpret generative outputs as executable toolpaths. Through this, we aimed to develop methods that translate spatial density and visual rhythm into continuous toolpaths that a robotic arm could perform.
How can generative images be translated into continuous fabrication toolpaths without losing their spatial intention? While the generative output used in this project forms an extendable, continuous pattern, it does not inherently encode the continuity required for robotic execution. For robotic arm 3D printing, the toolpath which it follows needs to be connected.
To address this, we structured the production process into three layers: a base layer for material grounding, a contour layer to define structural edges where the , and a layer for large-scale spatial organization. This layered approach allows us to examine which visual features of the generative image, such as density, depth, and regional contrast, carry structural potential when translated into motion.
My work focuses on interpreting the image through feature extraction. I use a depth map derived from the generative output to extract contour lines, treating depth as a proxy for spatial hierarchy rather than visual surface detail. This allows structurally significant regions to emerge independently of texture or noise. From these contours, I further design continuous infill patterns that connect isolated contours into executable toolpaths suitable for robotic arm fabrication.
The fabrication toolpaths are generated using three algorithmic pattern strategies. Each method interprets the same generative image through a different spatial logic, revealing how abstraction choices directly affect material behavior.
(1) Marching Squares
A traditional hatch pattern is adapted to follow the contours of the extracted lines. This method emphasizes linearity and continuity, allowing the robotic arm to maintain a steady motion while filling in areas between contours.
(2) Cross Hatching
Encodes tonal variation through directional stroke density, producing rhythmic, motion-driven material accumulation.
(3) Voronoi Subdivision
Introduces local variation through spatial partitioning, allowing regional differentiation within an overall system.
Across all three strategies, the same generative image produces radically different fabrication outcomes. Voronoi subdivision patterns create more expressive outcomes, as a result we choose it as the infill for the largest section.
Fabrication Reshapes Generative Logic
Material constraints, motion continuity, and irreversibility transformed visual density and spatial intention, exposing a fundamental mismatch between image-based generation and physical execution.
What generative qualities are worth preserving when execution demands continuity?
When execution demands continuity, generative value shifts away from surface detail and noise toward spatial intention—density gradients and regional coherence.