AI Teaches Harvesting Robots to Assemble Objects

2019-11-01

Manipulating objects of various shapes is not a strong suit for machines, but it is undeniable that this skill is highly useful for robots. In order to advance the latest technology in this field, researchers from Stanford University and Columbia University have recently studied a machine learning system called Form2Fit, which aims to teach pick-and-place robots with suction arms to assemble objects into complete toolkits. "If we combine the things that robots can learn, then perhaps they can adapt to new tasks better than before, which helps improve our efficiency in performing tasks such as assembling DIY furniture kits or packaging gifts into boxes," wrote research intern Kevin and robotics scientist Andy in a blog post. As they explained, Form2Fit primarily learns how to identify correspondences between objects through iterative experiments. One component, the two-stream matching algorithm, can derive a three-dimensional point representation that not only conveys the geometric shape of objects but also their textures. However, because the orientation of the point representation is sensitive, they have instilled in Form2Fit the knowledge of proper error elimination through rotation before placing objects in the target position. Two separate components of Form2Fit can generate effective picking and placing candidates: the attraction model and the planning model. The former can capture three-dimensional images of objects, while the latter prioritizes obtaining images of the target position, outputs successful placement predictions, and then integrates the outputs of the other three components of Form2Fit (including the matching algorithm) to generate the final picking position, placing position, and rotation angle. The team created a training dataset using the concept of time reversal disassembly, which means reversing the process and turning disassembly into assembly, and vice versa. Researchers used this dataset to train Form2Fit by randomly selecting a complete toolkit for disassembly and then reversing the disassembly procedure to learn how to assemble the toolkit. After 12 hours of continuous training, the researchers reported that the robot learned effective picking and placing strategies for various objects, achieving a success rate of 94% in different configurations of toolkits. When handling completely new objects and toolkits, the success rate exceeded 86%. "While the test results of Form2Fit are great, it still has limitations and cannot be widely applied at the moment," concluded the team's researchers. "Therefore, we will continue to conduct in-depth research to enhance the universality of Form2Fit and make it more practical."

RECENT AI NEWS

RECENT AI TOOLS