Google’s TossingBot Uses AI to Throw Objects Better Than Humans

Pick-and-place robots have been around for years and are able to perform their functions very well for a good reason — they were programmed to do so. Placing a specific object on a PCB or in a bin is remarkably easy to accomplish, as long as that robot can identify that object, but what if the objects are arbitrary, and the placement out of the range of the robot, could it still perform with precision? It is for Google’s TossingBot, at least 85% of the time anyway.

Google engineers designed the TossingBot with help from Princeton, Columbia, and MIT, which can learn how to pick up different objects and toss them into the right containers using machine learning. According to Andy Zeng at Google’s Brain Team, “TossingBot jointly learns grasping and throwing policies using an end-to-end neural network that maps from visual observations (RGB-D images) to control parameters for motion primitives. Using overhead cameras to track where objects land, TossingBot improves itself over time through self-supervision.”

As the video explains, the TossingBot uses a deep neural network and simple physics to learn how to toss randomly shaped and weighted objects into different positioned bins. The system employs three neural networks to accomplish that feat — one network identifies objects inside a container, a second to determine how to pick them up, and a third that tells the robot how to throw the object.

After 14 hours of trial and error, the TossingBot was able to toss the correct item into its designated target container about 85% of the time. The robot is capable of processing more than 500 objects per-hour, while typical pick-and-place platforms average 200 to 300 in the same timeframe.