Research on AI for CAD, 3D generative models, and engineering design agents. * denotes equal contribution. See also Google Scholar.
A state-conditioned LLM policy proposes CAD programs step by step, and an IoU-guided tree search refines them. Trained on ARCADE-1.5M, a new dataset of 1.5M executable CAD programs with sequences of 150+ steps, StepCAD improves IoU by up to 87.2% over the best prior method.
A hybrid optimization framework that recovers complex, editable CAD construction sequences (extrusions, revolutions, fillets, chamfers) from meshes by fitting and validating operations with IoU feedback. It sets the state of the art on DeepCAD, Fusion 360, and ABC, and also works end to end from photos.
Aligns SDF decoders in a shared weight space and generates new designs by parameter-constrained affine mixing. It enables controlled interpolation from as few as 50 samples, safe extrapolation up to 100% beyond the training ranges, and performance-driven shape optimization, such as minimizing aerodynamic drag on DrivAerNet++ cars and BlendedNet aircraft.
Encodes additive-manufacturing simulations as implicit neural fields whose weights follow the affine structure of the process parameters, predicting post-cooling distortion for unseen geometries, laser powers, and speeds.
The first large-scale dataset for UI interaction in precision engineering software: 41K+ annotated videos of CAD operations with time horizons up to 20× longer than existing UI datasets. It also introduces VideoCADFormer, a model that learns CAD interactions from video, and a VQA benchmark for 3D spatial reasoning in multimodal LLMs.
A unified benchmark of 18,000 samples across six families (DeepCAD, Fusion 360, ABC, MCB, Objaverse) and five input modalities. It evaluates eleven CAD-specialized and general-purpose VLMs over 1.4M+ generated CAD programs and identifies recurring failure modes.
MT-QWALE lets a robot arm complete Meta-World tasks in a single trial under distribution shift. It reaches near-100% success with randomized object positions and outperforms MT-SAC.
DQN agents that use segmentation and depth preprocessing to drive urban scenarios safely in the CARLA simulator, reaching a 94% success rate on unseen test trajectories.
Experimental characterization of the 3D-printed elastomer Tango Black under uniaxial tension, cyclic loading, and relaxation, measured with digital image correlation and fitted with a hyperelastic–viscoelastic constitutive model.