There are a lot of things that we humans take for granted, such as intuitively being able to follow directions with lots of complicated steps to them. An example is having someone help you put away dishes. If you ask a person to hang a coffee mug by the handle on a hook, they grab it and make the hanging happen without much thought.
The challenge for robots to perform similar tasks is that there are lots of other steps in that process that a robot has to learn to deal with. MIT says that the list of actions includes everything from finding the mug on the table to figuring out where to grab it, to figuring out where the handle is, and then locating the hook and putting the cup away. Other little steps are in between those steps.
Those steps can be tough for robots to handle. MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) has a new system for robots that allows the robot to perform a variety of pick-and-place tasks. The cool part is that the system enables the robot to perform the complex task of putting items away even if it has never seen the item before.
The researchers use a system called KPAM, short for Keypoint Affordance Manipulation. The system builds on previous research in using keypoints. KPAM collects a series of coordinates on an object. The coordinates provide all the data the robot needs to determine what to do with the object. In testing, the robot was able to place mugs accurately with only three keypoints for the center of the mug’s side, bottom, and handle.
For the shoe, the KPAM system needed six keypoints, and with that data, it could pick up over 20 different pairs of shoes running the gamut from slippers to boots. The system allows the incorporation of new object types quickly. The team says the bot was unable to pick up high-heeled shoes initially because there weren’t examples in the initial data set, adding a few pairs of heels to the data set allowed the robot to function.