Robot arms are getting smarter and smarter all the time, and now a team from Georgia Tech has taught them how to make their own tools. These resourceful robots are first taught to match form and function, then let loose on a series of everyday objects to build what they need to complete a given job.

They might not look as outright impressive as the back-flipping humanoids out of the Boston Dynamics labs, but robot arms are quietly developing some rather impressive intelligence. In recent years they've been taught to sort objects based on shape, learned how much force to use to pick up different foods, figured out how to handle objects they've never seen before, taught each other how to do their jobs and even been given a rudimentary self-awareness.

Now, the Georgia Tech team has taught them how to build tools using what's on hand. The project began in 2012 with the aim of making robots that could go into a disaster area, assess the situation and then help out using whatever it found lying around – so, for example, it would be able to use a loose pipe to push rubble off a survivor, or lay a door down as a bridge. This kind of creative thinking earned the planned robot the nickname "MacGyver bot," after the 80s TV character famous for inventive uses of everyday objects.

Seven years on, the MacGyver bot is learning to identify and use objects, although it still seems to be a long way off cruising into disaster zones. The robot is trained through machine learning techniques to understand what shapes let objects do certain things, and then put that knowledge to use to build a new tool. So, for example, it knows that concave objects like bowls can hold liquids, and use this knowledge to make a spoon.

The team also taught the robots how to attach things to make their new tools, either by piercing or grasping them. In the end, the robot arms managed to make hammers, spatulas, scoops, squeegees and screwdrivers.

"The screwdriver was particularly interesting because the robot combined pliers and a coin," says Lakshmi Nair, a researcher on the study. "It reasoned that the pliers were able to grasp something and said that the coin sort of matched the head of a screwdriver. Put them together, and it creates an effective tool."

The team says that at the moment, the robots can only "think" about the shape of objects and how to combine them. They can't yet take into account the properties of materials, such as how hard or soft they may be. But this is the next phase of the project.

"People reason that hammers are sturdy and strong, so you wouldn't make a hammer out of foam blocks," says Nair. "We want to reach that level of reasoning in our work, which is something we're working on now."

The research was presented at the 2019 Robotics: Science and Systems conference (PDF).

Source: Georgia Tech