Gilbreth 2.0: An Industrial Cloud Robotics Pick-and-Sort Application

Abstract

In prior work, we proposed an autonomous object pick-and-sort procedure for an industrial robotics application called Gilbreth. In this work, we developed improvements to two critical components of this application: object recognition and motion planning, integrated the new modules to create Gilbreth 2.0, and evaluated its performance. We used a Convolutional Neural Network (CNN) based object-recognition technique, which reduced object recognition time by a factor of 10 when compared to our previous solution, which used correspondence grouping. But this reduction in object recognition time came at a cost of requiring CNN model training time, which was 3 hours with just 13 object types. Our motion planning pipeline improvement was primarily to place constraints on the time threshold for each phase of the robot arm motion. This change enabled an improvement in the percentage of successful trajectories while keeping variance small. Finally, we evaluated the overall pick-and-sort performance of Gilbreth 2.0. We found that if the mean inter-object spawning time was 14 sec, while the mean service time for the robot arm to execute all phases of its motion was 13 sec, an overall pick-and-sort success rate of 71.3% could be achieved. We identified the causes of the failures, and found that further improvements are required to reduce motion-planning failure and grasping failure, and excess-load failure can be reduced further by increasing the inter-object spawning intervals.