[Summary: A new approach could cut learning time for robots that work with soft objects like ropes and fabrics. In simulations, the expanded training data set doubled the success rate of a robot looping a rope around an engine block. Using only the initial training data, the simulated robot hooked the rope around the engine block 48% of the time. After training on the augmented data set, the robot succeeded 70%.]
A new approach could cut learning time for robots that work with soft objects like ropes and fabrics.
It was developed by robotics scientists at the University of Michigan and might minimize the time it takes to learn new things and settings from a week or two to a few hours.
For their false data, they focused on the three features. It has to be credible, diverse, and relevant.
The simulated robot successfully wrapped the rope around the engine block 48% of the time using only the basic training data.
A new method enhances training data sets for robots that function with soft objects like ropes and fabrics, or in dense settings, advancing toward developing robots that can learn on the fly as people do.
Robotics scientists at the University of Michigan developed it and might minimize the time it takes to learn new things and settings from a week or two to a few hours.
In simulations, the larger training data set more than doubled the success rate of a robot looping a rope around an engine block and improved it by more than 40% from that of a physical robot executing the same task.
That is one of the activities a robot mechanic would need to gain enough skill at. However, according to Dmitry Berenson, U-M associate professor of robotics and senior author of a paper presented today at Robotics: Science and Systems in New York City, learning how to manipulate each unfamiliar hose or belt would require tremendous large amounts of data, likely collected for days or weeks.
During that period, the robot would experiment with the hose, extending it, connecting its ends, looping it around items, and so on, until it was aware of all the possible movements the hose could make.
Berenson said, “If the robot needs to play with the hose for a long time before being able to install it, that’s not going to work for many applications.”
Certainly, a robot colleague that needed that much time would probably not be well regarded by human mechanics. Thus, Berenson and Peter Mitrano, a robotics Ph.D. student, altered an optimization algorithm to assist computers to make some of the generalizations that people do, such as projecting how dynamics seen in one instance would recur in others.
On one occasion, the robot moved cylinders across a packed floor. The cylinder sometimes didn’t make contact with anything, but other times it did and the other cylinders shifted as a result.
You can replay the motion anywhere on the table where the trajectory does not drive it into other cylinders if the cylinder didn’t hit anything. A human would comprehend this, but a robot would need to acquire that data.
And Mitrano and Berenson’s program can also provide variations on the output of that first experiment that aid the robot in the same way rather than undertaking time-consuming experiments.
For their false data, they focused on the three features. It has to be credible, diverse, and relevant. Data on the floor, for instance, is meaningless if you’re focusing only on the robot moving the cylinders on the table.
On the other hand, the data must be varied; all parts of the table and all viewpoints must be evaluated.
“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,” Mitrano said, adding that both are important.
- How can we make sure that AI does what we suppose it to do?
- Scientists grew living human skin around a robotic finger
- Will our Future Look Like a Video Game?
The data must also be accurate. For illustration, any simulations in which two cylinders occupy the very same space would need to be labeled as invalid for the robot to know that this won’t happen.
Mitrano and Berenson broadened the data set for the rope simulation and experiment by extending the rope’s position to additional locations in a digital form of a real-world situation, assuming the rope would act the same as it did in the initial case.
The simulated robot successfully wrapped the rope around the engine block 48% of the time using only the basic training data. The robot has a 70% rate of success after training on the expanded data set.
Having the robot expand each effort in this way nearly doubles its success rate throughout 30 attempts, with 13 successful attempts as compared to seven, according to an experiment studying on-the-fly learning with a real robot.
This finding will be a big step for robotics research and human-robot interaction. This change will be helpful to the development of robots that can learn on the fly as humans do.
In addition, scientists have also predicted that this research will pave the way for a better understanding of how to use non-traditional methods in robotics such as data augmentation approaches to get better results.
The Toyota Research Institute, the Office of Naval Research, the National Science Foundation, and IIS-1750489 and IIS-2113401 funds all provided support for this research.