Artificially intelligent software can drive robots to perform the most menial tasks, such as reaching out and gripping objects.
However, there's one thing they can't, er, grasp easily. And that's dealing with things that move unexpectedly, which right now rules them out of a lot of real-world labor. However, it does make them good for picking up stuff that typically stays still, such as clothes on the floor or boxes in a warehouse.
It’s a surprisingly difficult skill to master, according to a paper to be presented at the Robotics: Science and Systems conference starting on Tuesday at Carnegie Mellon University in the US.
"We have been able to program robots, in very controlled environments, to pick up very specific items," Jürgen Leitner, a postdoctoral research fellow at the the Queensland University of Technology in Australia, said on Monday.
"However, one of the key shortcomings of current robotic grasping systems is the inability to quickly adapt to change, such as when an object gets moved."
Leitner and his colleagues built a neural network to help robots grasp objects that they haven’t been explicitly trained on. If they are to be useful in the real world, bots will encounter new objects and need to adapt to different environments.
"The world is not predictable – things change and move and get mixed up and, often, that happens without warning – so robots need to be able to adapt and work in very unstructured environments if we want them to be effective," Leitner added.
"For example, in the Amazon Picking Challenge, which our team won in 2017, our robot CartMan would look into a bin of objects, make a decision on where the best place was to grasp an object and then blindly go in to try to pick it up."
The neural net, a Generative Grasping Convolutional Neural Network or GG-CNN, differs from most traditional convolutional neural networks (CNNs) used in robotics. To grabs a thing, it has to determine how to position the robot's fingers to grasp an object. It does this using pixel-wise analysis of the input image instead of a more traditional sliding window or bounding box outlining the edges of the thing.
The GG-CNN model required fewer parameters than most CNNs so it’s much faster to execute, and takes about 19 milliseconds to run on a desktop computer equipped with a 3.6GHz Intel Core i7-7700 CPU and an Nvidia GeForce GTX 1070 graphics card.
"The Generative Grasping Convolutional Neural Network approach works by predicting the quality and pose of a two-fingered grasp at every pixel," said Douglas Morrison, first author of the paper and a PhD researcher at the Queensland uni.
"By mapping what is in front of it using a depth image in a single pass, the robot doesn't need to sample many different possible grasps before making a decision, avoiding long computing times."
It was trained by inspecting 885 images of real objects that have been labelled with positive points and negative points, which are where the robot should and shouldn’t aim to grip. Using a Kinova Mico 6DOF robot fitted with a Kinova KG-2 two-fingered gripper, it managed an 83 per cent grasping success rate on a series of eight 3D-printed objects with odd shapes, and 88 per cent on a range of 12 household items such as a screwdriver, teddy bear, and mug.
“This has benefits for industry – from warehouses for online shopping and sorting, through to fruit picking. It could also be applied in the home, as more intelligent robots are developed to not just vacuum or mop a floor, but also to pick items up and put them away," Leitner concluded. ®