Helping Robots Learn On The Fly With Augmented Data

Consumer Tech Helping Robots Learn On The Fly With Augmented Data Jennifer Kite-Powell Senior Contributor Opinions expressed by Forbes Contributors are their own. New! Follow this author to improve your content experience. Got it! Jul 21, 2022, 02:10pm EDT | New! Click on the conversation bubble to join the conversation Got it! Share to Facebook Share to Twitter Share to Linkedin Teaching robots to learn with augmented data.

getty A new study from researchers at the University of Michigan shows that feeding a robot fabricated or augmented data when no data sets were available increases the ability of that robot’s performance by 40%. According to a May 2022 study , Data augmentation for manipulation , the success of deep learning depends [. .

] on the availability of large datasets. Still, in robotic manipulation, there are learning problems with few or no datasets. According to the researchers, collecting the datasets is time-consuming and expensive, so one method to tackle the lack of data is data augmentation – creating additional training examples by modifying existing ones.

“For example, a robot hooks a rope around an engine block – that’s a task that any robot mechanic would need to be able to do,” said Dmitry Berenson , University of Michigan associate professor of robotics and senior author of the paper. “But learning how to manipulate an unfamiliar hose or belt would require huge amounts of data gathered for days or weeks. ” Getting on-the-job training is acceptable for a human, but for a robot, that can be time-consuming and costly to the manufacturing floor.

“To become accustomed to a new part or feature, [. . ] the robot would need to play around with the hose or rope—stretching it, bringing the ends together, looping it around obstacles [.

. ] until it understood all the ways the hose could move,” said Berenson. “If the robot needs to play with the hose for a while before installing it, that’s not optimal for applications in the real world today.

” MORE FOR YOU Google Issues Warning For 2 Billion Chrome Users Forget The MacBook Pro, Apple Has Bigger Plans Google Discounts Pixel 6, Nest & Pixel Buds In Limited-Time Sale Event Intuitive to humans, not to robots In the study, Berenson and fellow researcher Peter Mitrano, a doctoral student in robotics, decided to alter the optimization algorithm to enable a computer to make some of the generalizations humans make in predicting how dynamics observed in one instance might repeat in others. In one example, the robot pushed cylinders on a crowded surface. In some instances, the cylinder didn’t hit anything, but in other exercises, the cylinder collided with other cylinders, which then moved in response to that action.

Researchers said that if the cylinder didn’t bump into anything, that motion could be repeated anywhere on the table where the trajectory doesn’t make it crash into other cylinders. Sound obvious? Berenson says this action is intuitive to a human, but a robot will need that data to change its movement. So rather than doing a lot of time-consuming experiments, Berenson and Mitrano’s program created variations on the result from the first experiment that served the robot in the same way – creating fabricated data.

The team focused on three main qualities for their fabricated data – relevance, diversity and validity. “For instance, if you’re only concerned with the robot moving cylinders on the table, floor data isn’t relevant; the flip side is that data must be diverse. “If you maximize the diversity of the data, it won’t be relevant enough.

But if you maximize relevance, it won’t have enough diversity; both are important,” said Mitrano. The results? In the looping rope simulation, the researchers expanded the data set by extrapolating the rope’s position to other locations in a virtual version of physical space. The constant was that the rope would behave the same way it did in the initial simulation.

The simulated robot hooked the rope around the engine block 48% of the time using only the initial training data. But, after training on the augmented data set, the robot succeeded 70% of the time. Follow me on Twitter .

Jennifer Kite-Powell Editorial Standards Print Reprints & Permissions.