We have all had to clean up a bunch of small items at some point that we didn’t take the time to place into the correct bin and simply tossed the items into the bin from afar. Google is giving details on a new robot that it has developed that uses physics and deep learning to place items into appropriate containers. Google’s bot took first place in the stowing task at the Amazon Robotics Challenge.
It’s easy to see robots of this sort replacing humans at Amazon facilities in the future for picking items and packing them for shipment to buyers. Google built the robot with features that kinematically prevent it from dropping objects due to unforeseen dynamics. It uses steady and deliberate movements to keep an object from slipping.
The bot is also designed to tolerate the dynamics of an unstructured world. Google teamed with Princeton, Columbia, and MIT to develop TossingBot. The name comes from the design that is meant for the robot to learn to grasp and throw objects into selected boxes outside of its natural range. The bot learns by using an end-to-end neural network that maps from visual observations via RGB-D images to control parameters for motion primitives.
Overhead cameras track where objects land and self-supervision allows the robot to improve over time. TossingBot harnesses projectile ballistics to provide estimated for the throwing velocity needed to get the object to land at a target location. The neural network predicts adjustments required on top of that estimate to land at the correct location. The tech gives TossingBot throwing accuracy of 85%.