• Grasping
• Object manipulation
• Planning

Humans, thanks to their intelligence and physical skills, can easily manage an object which is partially covered so to make it easier to be grasped. Robots still lack that skill due to their perception, action capabilities and the complexity of reasoning about physical planning. Robotic planning approaches that try to bridge such gap have been studied in the last decades. One of the most promising is the rapidly exploring random trees. However, such an approach is usually limited to kinematic level planning due to the cost of managing large state-spaces and to the difficulty of modeling interaction forces. Recently, this application was extended to planning robotic manipulation in constrained environments. However, that approach is limited to manipulating objects shaped as boxes. This thesis generalizes that approach to objects of arbitrary shapes. The goal is achieved by implementing a new algorithm for the identification of the contact points between the object and the environment, and an algorithm for placing the fingers of the robot hand on the object even when this is partially occluded by the environment itself. The proposed algorithms rely on the application of the Gilbert–Johnson–Keerthi and of the voxelized hierarchical approximate convex decomposition algorithm.
The newly proposed method is tested numerically first, and then on a real-world scenario where a Franka robotic arm is used to manipulate and grasp several objects of increasingly complex shapes.