Sam Khalandovsky

Building on research done by Andreas ten Pas and Rob Platt, I worked on identification of object handles and automatic grasping of objects. The robot CARL was equipped with a 3-fingered 6-DOF JACO arm, which I used for the grasping tasks. Figuring out how to grasp an object is a difficult problem, and most existing solutions approach it in one of two ways: using a database of objects with precomputed grasps, or searching over the space of possible grasps for the given object geometry. Our approach was to simplify the problem to a search for cylinders and cylinder-like shapes designed for human grasping; once detected, forming a good grasp became a much easier problem. Most household objects are designed to be picked up by a cylindrical handle, or are themselves cylindrical, and cylindrical objects are relatively easy to grasp knowing only the position, orientation, and radius of the cylinder. Additionally, finding cylinders in a point cloud is a much simpler task than searching over all possible grasps. Thus, at a high level our grasping algorithm was as follows:

• Receive a pointcloud of the scene from the Kinect mounted on CARL
• Identify all cylinders that match radius and height parameters
• Group together close-by cylinders, representing each handle as a sequence of cylinders
• Present the user with a list of identified handles
• Calculate a trajectory to grasp the selected handle, based on the parameters of the cylinder closest to the handle's center