CADENA Rebuilds CAD Models Step by Step

Introduction

Converting an existing 3D shape into an editable CAD model remains a difficult engineering task. A mesh can capture the surface of an object, but it does not preserve the design history that engineers need: sketches, extrusions, cuts, and other parametric operations. In practice, reverse engineering a part often means rebuilding it feature by feature.

CADENA addresses this problem with a workflow that is closer to how human CAD users work. Instead of producing an entire CAD program in a single pass, the model grows the program step by step. After each predicted operation, the partial program is executed, and the resulting geometry is compared with the target mesh. The next operation is then conditioned on what has already been built and what is still missing.

Key points

• Iterative CAD reconstruction: CADENA frames reverse engineering as a sequence of executable modeling operations rather than a one-shot text generation problem.
• Geometry-in-the-loop feedback: Each intermediate result is checked against the target, giving the model a clearer signal about the remaining shape.
• Benchmark for real parts: The authors introduce CADENA-Bench, which contains 3396 deduplicated real mechanical parts from three industrial corpora.
• Per-family evaluation: The benchmark groups parts into six families and reports results by family, making method-specific weaknesses harder to hide behind aggregate averages.
• Released resources: The project provides inference code, supervised and reinforcement-learning weights, and the benchmark dataset.

Why it matters

The important shift in CADENA is not only better benchmark performance, but the move from static program prediction to a closed loop of prediction, execution, inspection, and correction. CAD models are not just shapes; they are editable construction histories. A system that can reason over intermediate geometry is better aligned with the way engineers create usable models.

According to the material, CADENA outperforms previous methods on CADENA-Bench as well as DeepCAD, Fusion 360, and MCB. It also reaches 0.910 voxel IoU on BenchCAD’s Vision2Code. The authors report that reinforcement learning against executed geometry improves accuracy and reduces the invalid rate to 0.9%, because operations that fail to build geometry receive no reward.

The benchmark may be just as important as the model. The authors note that learned baselines lose roughly half their score when moving from DeepCAD to real mechanical parts, highlighting a gap between controlled datasets and industrial geometry. For CAD AI, the next frontier is not simply generating plausible shapes, but producing reliable, executable, and editable CAD programs for real engineering workflows.

Source: Hugging Face Daily Papers