
85% success and 10% defects: why Amazon robots won’t replace humans yet
Amazon has tested its robots for warehouse logistics and came to an interesting conclusion. Machines are not yet ready to completely replace humans. Two robots participated in a series of experiments — Stow and Pick. Created for placing and selecting goods.
The Stow robot is equipped with a machine vision system, can evaluate free space in cells, and has a special grip with a retractable panel. Its task is to place goods in hanging fabric storage modules. During testing, it processed more than 500000 items. And handled about 85% of them. However, almost 10% of unsuccessful attempts led to damage to goods. Books fell to the floor or pages were crumpled during placement.
224 versus 243 units per hour — humans won. Despite the fact that Stow’s speed almost matched human speed, the amount of defects turned out to be too high for complete process automation.
The Pick robot showed better accuracy. 91% successful item selections over 6 months of testing. But it also has a drawback. In almost 20% of cases, it refused to perform the task, either by not recognizing the object or deciding there was a risk of damaging it.
And now Amazon will work not with manual algorithm tuning, but will train robots to “see and act” like a human.
It turns out they are betting on an “observation learning” approach, reminiscent of human learning. Which will potentially solve the problem of limited adaptability of robots when working with diverse goods.