This Robot Learns by Playing With Toys Like a Baby
In most cases, robots are programmed with specific instructions on how to interact with their surroundings. Even the ability to recognize a particular object has to be setup intentionally. Autonomous robots have some ability to improvise, but even that is limited to the system’s ability to categorize items according to known types.
As we’re seeing in other areas of artificial intelligence, the key to overcoming those challenges is to allow the robot to learn on its own. This may seem obvious when you consider the way young animals, including humans, learn by experimenting with the surrounding world. Researchers from the University of California Berkeley call this motor babbling, and they’ve now taught a robot to do the same thing.
The SCARA-style robot arm, called Vestri, starts by spending a week or so simply moving small objects around on a table. These are items like kids toys and other non-uniform objects you’d likely find around an average home. During this time, Vestri is learning how these objects respond to its touch, as well has how it itself moves.
Once the learning process has been completed, Vestri can then successfully move those objects to a desired location about 90% of the time. It accomplishes that by taking what it learned about physical interactions, and using that information to predict multiple potential scenarios for a request. It then chooses the prediction with the highest chance of success. Vestri is still in it’s early stages, but the research looks very promising for furthering AI actions.