A team at Google, in collaboration with individuals from Columbia, Princeton, and MIT Universities, have created a robot that can pick out objects from a bin and toss it into another specified bin.
The researchers said their biggest hurdle was having the robot grab an object and throw with precision. By grabbing random objects that have different edges, curves, and weights, the researchers said it was critical the robot make decisions prior to tossing the object into a bin. When throwing, the bot has to convert its decisions into an action that sends the object to the target. In response, the team built a robot arm that can perform these actions with accuracy comparable to a human arm.
The robot is pretty simplistic, where the robot can bend its arm into a box full objects, pick one up, and toss it into a box that’s been divided. The tricky part is the bot must hit the selected target.
The building process for the robotic arm consisted of helping it learn its job, programming it to scan the objects and boxes, and enabling it to sense the objects it needed to toss. Additionally, the bot used a deep-learning network to teach it the motions of tossing each object.
After the arm was created, the team let the robot toss objects into boxes 10,000 times without any help. The programmers initiated code that enforced the robot to empty the bin box back into its own box once all of the objects had been used.
In total, the researchers reported that the system was approximately 87 percent accurate in grabbing an object and 85 percent accurate in throwing it. The team tried to compare their abilities to the robot and found they were less accurate.