In their experiments, the researchers used eight designs, including a high-heeled shoe, a chess set, a toy car, and a coffee mug. The system samples enough values of the design parameters to offer a good approximation of all the available options, but that number varies from design to design. In some cases, it was only a few thousand samples, but in others it was hundreds of thousands. The researchers also developed some clever techniques to exploit similarities in design variations to compress the data, but the largest data set still took up 17 gigabytes of memory.
Finally, the system generates a user interface, a Web page that can be opened in an ordinary browser. The interface consists of a central window, which displays a 3D model of an object, and a group of sliders, which vary the parameters of the object’s design. The system automatically weeds out all the parameter values that lead to unprintable or unstable designs, so the sliders are restricted to valid designs.
Moving one of the sliders – changing the height of the shoe’s heel, say, or the width of the mug’s base – sweeps through visual depictions of the associated geometries, presenting in real time what would take hours to calculate with a CAD program. “The sample density is high enough that it looks continuous to the user,” Matusik says.
If, however, a particularly sharp-eyed user wanted a value for a parameter that fell between two of the samples stored in the database, the system can call up the CAD program, calculate the associated geometry, and then run tests on it. That might take several minutes, but at that point, the user will have a good idea of what the final design should look like.
“Autodesk has simplified versions of this project,” says Ryan Schmidt, a senior principal research scientist and head of the Design and Fabrication Group at Autodesk Research, the research arm of the leading manufacturer of CAD software. “We have a thing called Project Shapeshifter that is very similar to what a lot of other people are doing right now, which is making these geometry generators that have a parametric model you can explore. But they all have this common problem: that you can very easily make something that won’t work on your printer. What I thought was super-exciting about this work is that it can prevent you from designing something that isn’t going to print or that isn’t going to be strong enough once you’ve printed it.”
“As soon as I saw the [Siggraph] talk, I sent it to the people who work on Shapeshifter and said, ‘You guys should have this,'” he adds. “The sliders are an interface to what they did, but the underlying technology actually has a much broader application, I think, than just this consumer product for CAD novices.”