Contact Models in Robotic Grasp Learning
TL;DR
Generative contact models that infer grasps directly from depth images are reviewed, highlighting their role in grasp synthesis, generalization, and integration into dynamic and perception-driven manipulation pipelines.
Abstract
Abstract Autonomous robotic manipulation faces persistent challenges in dynamic, unstructured environments, where traditional physics-based models struggle with incomplete or inaccurate object and contact data. Machine learning has enabled adaptive manipulation skills, but purely end-to-end approaches often make contact-level reasoning implicit and rely on extensive training data. Structured methods that incorporate segmentation, depth, and planning offer complementary, more interpretable alternatives. Contact modelling plays a central role in predicting object motion and enabling realistic simulations, especially at close-range interactions. Early forward models demonstrated generalization to unseen objects, but struggled with action generation and exemplar dependence. Later advances introduced generative contact models that infer grasps directly from depth images, enabling one-shot learning and robust manipulation without prior object knowledge. This paper reviews and compares such contact modelling approaches, highlighting their role in grasp synthesis, generalization, and integration into dynamic and perception-driven manipulation pipelines.