For robots to interact effectively with objects in their environment, it is essential to obtain three-dimensional information about those objects -- specifically, their shapes and positions. Such information is typically obtained from data provided by...
For robots to interact effectively with objects in their environment, it is essential to obtain three-dimensional information about those objects -- specifically, their shapes and positions. Such information is typically obtained from data provided by vision sensors, most commonly RGB images, sometimes accompanied by depth information.
In recent years, fueled by rapid advances in machine learning, substantial progress has been made in processing visual data. Nevertheless, inferring accurate three-dimensional geometry and spatial configuration of objects solely from sparse and partially observed RGB images remains a highly challenging problem. Ensuring that machine learning–based perception methods function robustly and practically in real-world environments is an even greater challenge.
This thesis presents learning-based methodologies that, given sparse and partial RGB observations, can simultaneously identify individual objects in a scene and accurately reconstruct their three-dimensional shapes and positions. The proposed methods are specifically designed for perceptual robustness in real-world conditions, spanning shape primitive based approaches to generative modeling. Through extensive real-world manipulation experiments, this work demonstrates that the proposed methods retain strong performance beyond simulation. The central contribution of this thesis lies in developing robust 3D object shape perception methods from incomplete RGB observations that generalize reliably to real-world scenarios.
The first contribution is \textit{T$^2$SQNet} (Transparent Tableware SuperQuadric Network), a model that predicts low-dimensional geometric representations of transparent tableware objects using an extended deformable superquadrics. Leveraging the representational power of these primitives, the model captures the wide variety and complexity of real-world tableware geometries. Moreover, by relying solely on object mask images, the architecture minimizes the distribution gap between simulated and real-world data, thereby achieving strong real-world performance even when trained exclusively in simulation.
As a byproduct and contribution of independent interest, we also present {\it TablewareNet}, a publicly available toolkit for generating diverse datasets of transparent tableware with varying shapes and sizes, constructed using the proposed extended deformable superquadric representation. Experiments show that T$^2$SQNet, when trained on data generated by TablewareNet, outperforms existing methods for transparent object perception and proves effective in robotic manipulation tasks such as grasping and target retrieval.
Also, this thesis introduces DreamGrasp, a method that leverages pretrained generative image models trained on large-scale realistic datasets to infer occluded and unobserved regions of an environment. By combining coarse 3D reconstruction, contrastive-learning–based object separation, and text-guided per-object refinement, DreamGrasp overcomes the limitations of existing approaches and enables robust 3D reconstruction in cluttered, multi-object environments. Experimental results demonstrate that DreamGrasp not only reconstructs accurate object geometries but also achieves high success rates in sequential grasping and target retrieval tasks.