In a step toward robots that can learn on the fly like humans do, a new approach expands training datasets for robots that work with soft objects like ropes and fabrics, or in cluttered environments.
Developed by robotics researchers at the University of Michigan, it could cut the time it takes to learn new materials and environments to hours rather than a week or two.
In simulations, the expanded training dataset improved the success rate of a robot wrapping a rope around an engine block by more than 40% and nearly doubled the success rate of a physical robot for a task. similar.
This task is one that a robot mechanic should be able to accomplish easily. But using today’s methods, learning how to manipulate every unknown pipe or belt would require huge amounts of data, likely collected over days or weeks, says Dmitry Berenson, UM associate professor of robotics and author lead of a paper presented today at Robotics: Science and Systems in New York.
Meanwhile, the robot played with the pipe — stretching it, bringing the ends together, wrapping it around obstacles and so on — until it figured out all the ways the pipe could move. .
“If the robot has to play with the pipe for a long time before it can install it, it won’t work for many applications,” Berenson said.
Indeed, human mechanics probably wouldn’t be impressed with a fellow robot who needed that kind of time. Berenson and Peter Mitrano, a doctoral student in robotics, therefore modified an optimization algorithm to allow a computer to make some of the generalizations that we humans make – predicting how the dynamics observed in one case might be repeated in others. others.
In one example, the robot pushed cylinders across a cluttered surface. In some cases the cylinder didn’t hit anything, while in others it collided with other cylinders and they moved in response.
If the cylinder didn’t hit anything, this move can be repeated anywhere on the table where the trajectory doesn’t take it into other cylinders. It’s intuitive for a human, but a robot needs to get that data. And rather than doing time-consuming experiments, Mitrano and Berenson’s program can create variations on the result of that first experiment that serve the robot in the same way.
They focused on three qualities for their fabricated data. It had to be relevant, diverse and valid. For example, if you are only concerned with the moving cylinders of the robot on the table, the ground data is irrelevant. The flip side is that the data should be diverse — all parts of the table, all angles should be explored.
“If you maximize the diversity of data, it won’t be relevant enough. But if you maximize relevance, they won’t have enough diversity,” Mitrano said. “Both are important. »
And finally, the data must be valid. For example, all simulations that have two cylinders occupying the same space would be invalid and should be identified as invalid so the robot knows this will not happen.
For the simulation and the rope experiment, Mitrano and Berenson extended the dataset by extrapolating the position of the rope to other locations in a virtual version of a physical space – as long as the rope would behave like the same way as in the first example. Using only the initial training data, the simulated robot hooked the rope around the motor block 48% of the time. After training on the augmented data set, the robot was successful 70% of the time.
An experiment exploring on-the-fly learning with a real robot suggested that allowing the robot to extend each attempt in this way almost doubles its success rate over the course of 30 attempts, with 13 successful attempts instead of seven.
This work was supported by National Science Foundation grants IIS-1750489 and IIS-2113401, Office of Naval Research grant N00014-21-1-2118, and the Toyota Research Institute.
#extend #training #datasets #manipulation #tasks #improves #robot #performance #Artificial #Intelligence #Robotics #News