Robotic learning has quickly become of automation’s most vibrant categories — and understandably so. Programming a robot has traditionally required a lot of technical know-how, but what if there was a simpler way for non-programmers/roboticists to teach these systems to do what we want?
Imitation and reinforcement learning are two of the most popular methods at the moment. The first involves taking control of the robot to teach it to perform a task, while the second entails training a system on millions of images.
A number of researchers are exploring an even more intuitive method that effectively trains a system by watching a human complete a task. A team at Carnegie Mellon University is demonstrating the in-the-Wild Human Imitating Robot Learning, or WHIRL, an algorithm that can train a system by watching a video.
In their demos, an off-the-shelf mobile robotic arm learns to complete 20+ household chores, including opening and closing drawers and appliances and taking out the trash.
“Imitation is a great way to learn,” Robotics Institute PhD student Shikhar Bahl said in a release. “Having robots actually learn from directly watching humans remains an unsolved problem in the field, but this work takes a significant step in enabling that ability.”
It’s easy to see how a feature like this might prove particularly handy in a home setting, where roboticists anticipate these systems will one day be deployed to help elderly homeowners and other people with impaired movements.
In the case of WHIRL, no special adds-ons are required. The robot simply attempts to execute a certain task until it’s successful, even if it takes several times to fully master it. As CMU notes, its own approach may not be identical to the humans — instead, the system is looking to find the best method to complete the task based on its own hardware limitations.
Right now the system is trained by watching videos, and the team is looking to expand things out to include clips from services like YouTube.
