Robust Motion Generation for Vision-Guided Robot Bin-Picking

This paper discusses work towards a vision-based solution to the problem of robot bin-picking. The problem of robot bin-picking is defined as searching for and recognizing a part among many lying jumbled in a bin such that the robot is able to grasp and manipulate the part. Despite decades of research in vision, robotics, and manufacturing, this problem remains open. Currently, in modern manufacturing, this seemingly simple task is performed by complex assembly lines or manual labor. The amount of efforts and costs associated with the current solutions to bin-picking is a testament to the importance of a new solution. The main objective of this research is a reliable and cost effective automated solution to the bin-picking problem encountered in manufacturing. As a broader contribution, this research also provides a robust visual servoing method that enables safe interactions between a robot and its environment. Our system uses visual feedback to generate tasks autonomously and to control the interaction of the manipulator with its environment. First, our system relies on robust vision-based object localization to generate three-dimensional pose hypotheses for each identified part. Then, the hypotheses are filtered according to the feasibility of their picking configuration. Finally, a trajectory is generated for a picking position. In this paper, we consider the specifications of the trajectory ensure that collisions with the bin and joints limits are avoided, while servoing the robot to the part. To ensure the reliability of the system, the procedure is tested in a simulation before being executed by a manipulator. Our experiments target the automotive industry and involve real engine parts a typical industrial robot and metal bin.