Photo of Google's office and campus bicycles

TossingBot: New Google robotics division teaches robot to learn for itself

The robot uses machine learning and trial-and-error to teach itself how to pick up and accurately throw any object

Like GearBrain on Facebook

Nearly two years after selling off its Boston Dynamics robots division to Softbank, Google has got back into robotics with a new project.

Called the TossingBot, the robot, which comes from a new division called Robotics at Google, can teach itself through trial-and-error to pick up and accurately throw any object.

Read More:

While simple for humans, this action, Google explains in a blog post, is very difficult for a computer to understand. For example, humans quickly discover how the trajectory of a screwdriver will differ if you throw it while holding the handle, or the metal tip.

We understand that the way the object is weighted affects how it flies through the air, but robots powered by artificial intelligence need to learn this through trial-and-error. "Manually designing a solution that explicitly handles these factors for every random object is nearly impossible," Google says.

However, through deep learning, the company says it has created a robot which "can learn from experience rather than rely on manual case-by-case engineering."

Image showing a Google-built robot throwing an object into a boxThe robot is seen here accurately throwing an object into a boxPrinceton

Turning to its new robot, Google adds: "A fundamental component of TossingBot is that it learns to throw by integrating simple physics and deep learning, which enables it to train quickly and generalize to new scenarios."

The robot, created in partnership with researchers from Princeton, Columbia and MIT, was initially faced with random objects of various shape and weight. It repeatedly tried, and often failed, to grasp the objects, but after time it learnt what to do and become better at picking them up, which simultaneously improves its ability to throw them accurately at a target. Google adds: "Occasionally the robot randomly explores what happens if it throws an object at a velocity it hasn't tried before."

After 10,000 grasps and throw attempts (which Google says took 14 hours of training time), the robot was capable of picking up and throwing a random object with 85 percent accuracy. The robot was also able to pick up a random object with 87 percent reliability.

News of the Google robot comes just a few days after MIT showed off a robot of its own which uses machine learning to pick up and move objects it has never seen before.

We can see such a robot being useful at sorting through garbage and other waste materials picking out items which can be recycled - like aluminum cans, cardboard boxes and glass bottles - and throwing them into a receptacle. The technology may also prove useful in a retail fulfilment center, where many different objects are picked up and moved before being boxed up and mailed to customers.

Once it had got to grips with simple objects like wooden blocks, the robot then taught itself how to pick up and accurately throw more complex items, like fake fruit, decorations, and office objects. Within a few hundred attempts (taking an hour or two), the robot is able to accurately throw a new and complex object, Google says.

The robot does not yet understand what damage might occur to thrown objects when they land, and how this would differ from one to another, but Google says this extra degree of intelligence is on the TossingBot's roadmap. The company is also considering teaching other robots to catch thrown objects in a bid to soften their landing.

Like GearBrain on Facebook
The Conversation (0)

GearBrain Compatibility Find Engine

A pioneering recommendation platform where you can research, discover, buy, and learn how to connect and optimize smart devices.

Join our community! Ask and answer questions about smart devices and save yours in My Gear.

Top Stories

Weekly Deals