Quantcast

Engineers Program Robot To Learn From A Human’s Touch

November 6, 2013
Image Caption: Cornell engineers program a Baxter robot from Rethink Robotics to learn from humans. Credit: Cornell University

[ Watch the video: Human Touch Makes Robots Smarter ]

Lee Rannals for redOrbit.com – Your Universe Online

Industrial robots have no brains and are programmed to move through actions that have been embedded in the memory bank. These robots repeat the exact same action every time an object goes by, making them efficient but limited in capability.

Now, engineers have built a robot that has the ability to learn from humans and make adjustments to its memory bank while performing a task.

A team from Cornell University modified a Baxter robot from Rethink Robotics, giving it the power to co-actively learn from humans in order to plan actions that are suitable to the environment it is working in. This will enable robots working in a more diverse situation, such as in a house, to handle more objects.

A personal robot working in a house may have to handle both tomatoes and canned goods, each of which require a different grip. A co-active learning technique would help personal robots continue to learn from humans so that they can be useful in a variety of environments.

Researchers built upon previous work by adding programming that lets the robot plan its own motions. The modified Baxter robot displays three possible trajectories on a touch screen where an operator is able to select the one that looks the best.

As the robot executes its movements, the operator is able to intervene and manually guide the arms to fine-tune the trajectory. This technique enables operators to give the robot corrective feedback, effectively sharpening the machine’s talents.

The team designed a “zero-G” mode on the robot where its arms hold their position against gravity but allow the operator to move them. The robot has a learning algorithm that can memorize the movements the operator makes with its arm. While the first correction may not be the best option, the learning algorithm allows the robot to learn incrementally so it could keep refining its trajectory a little more each time the human operator makes adjustments or selects a trajectory on the touch screen.

Researchers said that it asked a group of people who were not part of the research team to attempt to train the robot. Most of the users were able to train the robot successfully on a particular task with just five uses of corrective feedback.

The Baxter robot is able to associate a particular trajectory with each type of object. For example, a quick flip over might be the fastest way to move a cereal box, but that would not work with a carton of eggs or an open bottle of liquid. The robot can also be taught certain aspects about objects, such as the fact that eggs are fragile and should be handled with care by keeping them close to the counter.

The robot can be taught self-awareness as well, so if it is cutting an object with a knife in the kitchen, it knows to hold the knife in close and not to make a wide swing action towards a human.

Researchers will soon be presenting their paper at the Neural Information Processing Systems conference in Lake Tahoe, California.


Source: Lee Rannals for redOrbit.com - Your Universe Online



comments powered by Disqus