Robots are learning to do many tasks for humans, but one task that is difficult for them to learn is how to catch an object. While we don’t think about everything that occurs in that one second when a fast-flying object is tossed to us, there are many things that a robot would have to learn to successfully catch it.
To be able to teach robots how to catch an object in-flight, the following three problems need solutions:
- An accurate prediction of the trajectory of the object, depending on the object’s shape
- Prediction of the optimal interception point
- Fast and precise planning of the robot’s arm to move at the correct moment to catch the object
While robots can be constructed with sensors, an arbitrary motion, like catching an object is very hard to sense. The estimation of the object’s flight and the location where it will land have to be calculated in a split second.
Researchers have developed a framework to teach this action to robots using rigid, uneven mass distributions and non-rigid mass distributions. Robots are taught either through observing humans or through exploration. Using modules to estimate the trajectory of the flying objects, determine the final catching configuration, and for the correct arm and finger motions, robots were able to learn how to catch.
Written by IEEE on August 30, 2017