An autoregressive framework for reconstructing editable parametric computer-aided design models from point clouds

Reverse engineering of industrial parts from three-dimensional (3D) point clouds is vital for modern computer-aided design (CAD) workflows in manufacturing and design. However, the unordered nature of point clouds and its substantial modality gap from sequential, parameterized CAD modeling make accurate and editable reconstruction particularly challenging. Existing methods often struggle to address both issues simultaneously, resulting in limited shape fidelity and practical usability. This work introduces an end-to-end autoregressive transformer-based framework termed RenCAD, which directly maps point cloud representations to CAD modeling sequences, enabling automated, parametric reconstruction. The primary artificial intelligence (AI) contribution of RenCAD lies in its overall autoregressive architecture, which generates CAD sequences token by token to capture inter-command dependencies. This is further enhanced by a group-based tokenization mechanism with learnable positional encoding to effectively handle unordered point clouds, and a proportional reconstruction loss that improves parameter learning with respect to geometric accuracy. From an engineering perspective, RenCAD reduces manual modeling effort while preserving editability and precision. It outputs parametric operations fully compatible with standard CAD software, supporting seamless integration into industrial workflows. Extensive experiments on a public benchmark demonstrate that RenCAD outperforms similar methods across multiple evaluation metrics, achieving higher command accuracy, lower invalid rates, and improved geometric fidelity. These results confirm the robustness and scalability of RenCAD, advancing AI-driven CAD automation by tightly coupling raw point cloud data with structured modeling logic. Nevertheless, RenCAD only supports a limited set of modeling operations. Future work will extend its modeling capabilities and adaptability to broader industrial scenarios.