Clustered Grasp Volumes for Improved Grasp Selection

Successful manipulation of objects in unstructured scenes requires identifying an achievable grasp given current environmental conditions. Selecting this grasp is a challenging problem, made more difficult when operating in adversarial conditions. Existing approaches often treat the grasp selection and subsequent path and motion planning as separate optimization tasks which require assuming a static environment. We propose a grasp selection algorithm that clusters grasps by pose and cost to generate discrete graspable volumes. Grasps are chosen from within each cluster by a cost function to maximize visibility of the target and manipulability of the robotic arm. This optimization framework is flexible and allows for autonomous or human-in-the-loop supervision in achieving a grasp through a closed-loop control process. Our approach is evaluated in simulation and in a test tank with a Reach Robotics Bravo 7 robotic arm. Code is available at https://gitlab.com/apl-ocean-engineering/raven/manipulation.