Researchers at Georgia Tech’s “Robot Autonomy and Interactive Learning (RAIL)” successfully trained a robot to create its own basic tools by combining objects.
This development is an important step towards more ‘intelligent’ machines capable of producing advanced tools that could be useful in hazardous environments.
The whole “philosophy” is reminiscent of television’s “MacGyver” the hero of the popular 1980’s television series, who solved problems and encountered difficult situations by improvising and combining / using simple objects at his disposal.
For many years, computer scientists and other researchers have been trying to give robots respective capabilities. In this particular case, RAIL researchers under Associate Professor Sonia Chernova used a technique developed by Mike Stilman, a former Georgia Tech professor.
On this case, the robot that was ‘trained’ through the method developed by the researchers was given a set of objects and was asked to make a specific tool. Like humans, the robot first looks at the shapes of objects and how they could be connected to each other.
Using machine learning, the robot trained to match shape with function -what shapes lead to concrete results- from a wide range of examples of everyday objects: For example, ”learning” how the shape of the bowl allows them to keep liquids, using this knowledge to create spoons. Respectively, the robot «trained» how to associate objects through examples with materials that can be perforated or gripped.
Within the framework of the research successfully created hammers, spatulas, screwdrivers and so on.
“The screwdriver was very interesting because the robot combined a plier and a coin,”
said Lakshmi Nair, a doctoral student in the School of Interactive Computing and one of the program’s researchers.
“He thought that the pliers were able to catch something, and that the coin somehow resembled a screwdriver head. Put them together and create an effective tool.”
At present, the robot is limited only to shape and connection; it cannot draw conclusions about specific properties of materials, which would be a crucial step in real-world application of this technology.
“People think hammers are tough and durable, so you wouldn’t make a foam hammer. We want to reach this level of inference, and we are working on it now, “