Robotic Grasp Planning with Foundational Models

This thesis explores the use of foundation models in the context of robotic grasp planning, addressing the challenge of achieving both geometric robustness and task-oriented adaptability in automated manipulation. The research focuses on the integration of two state-of-the-art frameworks: GraspGen, a diffusion-based 6-DoF grasp generation model capable of producing diverse and physically consistent grasp configurations, and FoundationGrasp which leverages large-scale vision-language representations for task-aware grasp evaluation. The proposed work aims to combine the semantic understanding of FoundationGrasp with the generative capabilities of GraspGen, developing a hybrid planning architecture implemented within the ROS 2 ecosystem. The system is designed to generate grasp candidates through probabilistic sampling and subsequently refine them using task compatibility scores derived from multimodal foundation models. YOLO was employed for perception (detection and segmentation), behavior trees for control logic, and Docker for containerization ensuring reproducible deployments across development and testing phases. The experimental evaluation demonstrates the benefits of this integration, showing improved task consistency and generalization across unseen objects and manipulation scenarios. This work contributes to the development of intelligent grasp planning systems that bridge the gap between low-level geometric reasoning and high-level semantic understanding, paving the way toward more generalizable and adaptive robotic manipulation.