Expanding Datasets for Robots

With traditional techniques, training robots often requires hundreds of hours of data, but this is not a practical way to train robots on every variation of a task. U-M researchers used data augmentation to develop a method that will expand these datasets. With a small amount of data, the researchers explored the task of a robot hooking a rope under an engine. In a virtual space, they held the rope in the same position and moved it around the scene to create copies of each simulation. They took this augmented data and applied it to a virtual and real robot, finding that the robots successfully completed the task with augmented data more times than without. This method will drastically cut down learning time for robots and move them a step closer to learning quickly like humans.

This research was led by PhD Student Peter Mitrano and Dmitry Berenson, Associate Professor at the the University of Michigan Electrical Engineering and Computer Science department and Robotics Institute. They are both a part of the Autonomous Robotic Manipulation Lab.

Read more:

------

The University of Michigan College of Engineering is one of the world’s top engineering schools. Michigan Engineering is home to 12 highly-ranked departments, and its research budget is among the largest of any public university.

Follow University of Michigan Engineering:

Contact University of Michigan Engineering: