Robotic bin picking is an active research topic which aims at picking up objects with random poses one by one from an unstructured clutter. It benefits a range of problems in human life, varying from industrial manufacturing to domestic applications. Despite its practicality, bin picking is still an open and challenging problem. Recent solutions to bin picking usually result in a direct pinch grasp on the object, where any other physical interaction between the robot gripper and the object is avoided. However, proper contact interaction is essential to successful singulation and simultaneously picking of the objects, especially those thin objects with large width-to-thickness ratio. Moreover, the gripper design that can facilitate the interaction is another important issue to investigat...[ Read more ]

Robotic bin picking is an active research topic which aims at picking up objects with random poses one by one from an unstructured clutter. It benefits a range of problems in human life, varying from industrial manufacturing to domestic applications. Despite its practicality, bin picking is still an open and challenging problem. Recent solutions to bin picking usually result in a direct pinch grasp on the object, where any other physical interaction between the robot gripper and the object is avoided. However, proper contact interaction is essential to successful singulation and simultaneously picking of the objects, especially those thin objects with large width-to-thickness ratio. Moreover, the gripper design that can facilitate the interaction is another important issue to investigate. In this thesis, we proposed model-based and data-driven approaches to develop manipulation techniques that can effectively singulate and pick thin objects from clutter, which is the goal of bin picking.

First, we present a new manipulation technique we called dig-grasping, for object singulation and picking. We take advantage of quasistatic planar pushing mechanism that models the physical interaction between the gripper and the object to pick during grasping. A gripper designed with digit asymmetry, that is, a length difference between two fingers, is suggested as a key to capture the object. We validate the effectiveness of this approach in 3D bin picking tasks through an extensive experiments. More complex tasks beyond picking, such as pick-and-place/pack are also demonstrated.

Second, we propose a data-driven approach where the robot learns to pick by digging. Dig-grasp manipulation shows the importance of appropriate physical interaction to successful bin picking, however, the model-based approach relies on accurate simulation of the object behavior and precise estimation of the object pose, which can involve large uncertainty and noise in reality. Hence in this work, we aim at constructing an end-to-end learning framework that learns the interactive picking action directly from visual information. Consequently, the robot is enabled to interact with the object clutter by performing a learned digging operation. We leverage the fully convolutional network (FCN) to predict the grasp success probabilities for a set of interactive action candidates, where the optimal action primitive will be specified to execute. The network is trained in a simulation environment through iterated self-supervision. We perform a set of bin picking experiments to verify the effectiveness and generalizability of the presented approach.

Finally, we develop a manipulation technique called tilt-and-pivot, to grasp the thin-layer object lying on a flat surface , and apply it to bin picking. This technique features in-hand manipulation where the relative configuration between the object and the gripper changes during the motion. We present the kinematics and planning of tilt-and-pivot, hardware design, and the effectiveness of the approach. We also present a set of experiments that show the applicability of the technique in bin picking scenario.