CADIR: A Cross-Backend Editable Intermediate Representation for Agentic CAD Generation
This paper introduces CADIR, a cross-backend executable intermediate representation built on the OCCT kernel that enables agentic CAD generation with explicit construction graphs and geometric signature matching to achieve high-fidelity, editable model reconstruction across diverse CAD systems like FreeCAD, SolidWorks, and Fusion 360.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are trying to teach a robot to build a complex Lego castle. You could give the robot a photo of the finished castle and say, "Make this," but the robot might just glue the bricks together in a messy pile that falls apart if you try to move a single piece. Or, you could give the robot a long, rigid list of instructions like "Pick up red brick 4, place it on blue brick 2," but if the robot's tools are slightly different from the ones you used to write the list, it might grab the wrong brick or get stuck. This is the current challenge in a field called Computer-Aided Design (CAD), where engineers use software to design everything from car engines to smartphone cases. For a long time, computers could only make static 3D shapes or follow very specific, rigid scripts that broke easily if you tried to change them or move them to a different software program. But now, with the rise of "Large Language Models" (AI that understands human language), we are starting to see computers that can listen to a description like "make a gear" and try to build it. The big question is: how do we make sure the computer builds something that is not only the right shape but also easy to fix, change, and use in different design programs?
This is where a new system called CADIR comes in, acting like a super-smart translator and project manager for AI builders. The researchers behind this paper found that existing methods were like giving a chef a recipe written in a secret code that only one specific kitchen could read; if you tried to cook it in a different kitchen, the ingredients wouldn't match, or the steps would make no sense. They also noticed that previous AI attempts often built "static" models—like a 3D statue—that you couldn't tweak later without starting over. To solve this, the team created CADIR, which stands for a "Cross-Backend Editable Intermediate Representation." Think of it as a universal, step-by-step construction log that doesn't just record what was built, but how it was built, keeping track of every dependency (like how a wheel is attached to an axle) in a way that any major design software can understand.
Instead of just spitting out a final 3D shape, the CADIR system builds a "construction graph." Imagine this as a dynamic flowchart where every step of the building process is a node, and the lines connecting them show how one step depends on another. If you change the size of a hole in the middle of the design, the graph knows to update everything connected to that hole, just like a real engineering blueprint. The system uses a clever trick called "Geometric Signature Matching" to find the right parts. It's like if you lost a specific Lego brick in a different set; instead of looking for a brick with the same serial number (which might not exist in the new set), the system looks for a brick with the same shape, size, and color, ensuring the new design fits perfectly even if the software is different.
The researchers tested this by asking the AI to build hundreds of mechanical parts, from simple blocks to complex assemblies with gears and moving parts. They found that CADIR was incredibly reliable, achieving a perfect "execution success rate" of 1.0000, meaning the AI successfully built the models without crashing or making errors every single time. When they compared it to other methods, CADIR produced shapes that were much closer to the intended design (improving accuracy by about 30% in some tests) and, crucially, allowed the models to be opened and edited in popular software like FreeCAD, SolidWorks, and Fusion 360. In fact, when they tried to edit these reconstructed models, the system succeeded in fixing or changing the parts nearly 100% of the time in FreeCAD and Fusion 360.
What makes this particularly exciting is that the system doesn't just build from scratch; it can also "remember" past designs. The researchers gave the AI a library of previous construction graphs and found that when the AI searched for similar shapes or parts using text or images, it could find the right building blocks much faster and more accurately than if it were just searching through raw code. This means the AI can learn from its own history, reusing clever little structures (like a specific way to attach a handle) to build more complex things later. While the paper notes that some very specific features in certain software still can't be perfectly edited due to the unique limitations of those programs, the overall result is a massive leap forward. It suggests that we are moving toward a future where AI can not only design complex machines but also hand them over to human engineers in a format that is ready to be tweaked, tested, and improved, bridging the gap between a simple idea and a real, working product.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.