Waste-sorting robots that have been learning their job while wandering through Google offices can now effectively sort items in bins into recycling, compost and rubbish.
One way of teaching machines to perform tasks is by reinforcement learning, in which a robot is told what a successful outcome looks like and is left to figure out how to achieve it by trial and error and a system of feedback, gradually building up an optimal model of what to do.
Advertisement
Normally, a robot starts to learn how to complete a task in this way from scratch, but to kick-start this process for trash-sorting robots, at Google and his colleagues used simulations and a classroom consisting of 20 robots at workstations where they could practise and improve their rubbish sorting in a controlled environment.
At the same time, 23 robots were placed in Google offices and left to find waste stations of unsorted or incorrectly sorted rubbish and rearrange the items they found so they were in the right place.
All the robots shared the same model of how to sort the waste, and that model improves the longer the robots operate, whether in simulations, classrooms or real-world sorting.
鈥淭he more robots that are running this model and running reinforcement learning, the better all of them get together,鈥� says Levine. 鈥淚n the experiments we did, we had on the order of tens of robots, but if there was some kind of actual real commercial application with thousands of robots out there then, in principle, the learning rate of all of them will be a lot faster.鈥�
Levine and his team let the robots sort people鈥檚 rubbish for nearly 10,000 hours over two years and, after that period, the machines managed, on average, to sort 84 per cent of items accurately in a classroom test. The robots had also learned their own intuitive behaviours for some waste, such as nudging larger objects over the bin鈥檚 edge or picking up multiple small items at the same time, says Levine. The researchers have posted details of .
鈥淭he system would surely be outperformed by trained professionals, but it seems to be correcting mistakes that we all make in our daily lives and busy schedules,鈥� says at the University of Sheffield, UK. He says these types of systems could provide real-world benefits in offices over the next five years.
Giving the robots the ability to detect what things feel like through tactile sensors might make them even better at sorting rubbish, suggests Gross.
