Robot dexterity: from deformable grasping to impulsive manipulation
Nowadays, it is fairly common for robots to manipulate different objects and perform sophisticated tasks. They lift up massive hard and soft objects, plan the motion with specific speed, and repeat complex tasks with high precision. However, without carefully control, even the most sophisticated robots would not be able to achieve a simple task.
Robot grasping of deformable objects is an under-researched area. The difficulty comes from both mechanics and computation. First, deformation caused by grasping motions changes the global geometry of the object. Second, different from rigid body grasping whose torques are invariant, the torques exerted by the grasping fingers vary during the deformation.
Collision is a common phenomenon in robot manipulation that takes place when objects collide together, as observed in the games of marbles, billiards, and bowling. To make the robot purposefully make use of
impact to perform better at certain tasks, a general and computationally efficient model is needed for predicting the outcome of impact. And also, tasks to alter the trajectory of a flying object are also common in our daily life, like batting a baseball, playing ping-pong ball. A good motion planning strategy based on impact is necessary for the robots to accomplish these tasks.
The thesis investigates problems of deformable grasping and impact-based manipulation on rigid bodies. The work contains deformable grasping on 2D and 3D soft objects, multi-body collision modeling, and motion planning of batting a flying object.
In the first part of the thesis, in 2D space an algorithm is proposed to characterize the best resistance by a grasp to an adversary finger which minimizes the work done by the grasping fingers. An optimization scheme is offered to handle the general case of frictional segment contact. And also, an efficient squeeze-and-test strategy is introduced for a two-finger robot hand to grasp and lift a 3D deformable object resting on the plane.
Next, an $n$-body impulse-based collision model that works with or without friction is studied. The model could be used to determine the post-collision motions of any number of objects engaged in the collision. Making use of the impact model, the final part of the thesis investigated the task of batting a flying object with a manipulator. First, motion planning of the task in 2D space is studied. In the frictionless case, a closed-form solution is analyzed, simulated, and validated via the task of a WAM Arm batting a hexagonal object. In the frictional case, contact friction introduces a continuum of solutions, from which we select the one that expends the minimum kinetic energy of the manipulator. Next, analyses and results are generalized to 3D. Without friction the problem ends up with one-dimensional set of solution, from which optimum is obtained. For frictional case hitting normal is fixed for simplicity. The system is then transferred to a root-finding problem, and Newton's method is applied to find the optimal planning.