Ding dong! That鈥檒l be the robot with my pizza. Such a scenario probably seems a bit far-fetched but, in the US and UK, delivery firms like JustEat and DoorDash are already experimenting using small robots to deliver groceries and meals.
Currently聽these systems need human chaperones to monitor the robot鈥檚 progress, jumping in if it gets into trouble. But now聽Kiwi, a company based聽at the University of California, Berkeley, is using machine learning to聽teach聽its delivery robots how to cross the road safely, without any human intervention. It could be an important step in making these robots more autonomous, something that is聽vital聽if they are ever going to be delivering our dinners at scale.
Such a system could also help delivery firms with the tricky 鈥榣ast mile鈥� problem of logistics 鈥� the fact that getting parcels to your door is the most expensive bit of the delivery process.
Advertisement
Kiwi and lets students order food from campus restaurants via an app, to be delivered聽by its small fleet of robots. The robots use a mixture of camera sensors, lasers and an in-built map of the campus to find their way between restaurants and student addresses.
Bots are doing it for themselves
Human operators monitor the robot at all times, making sure they are safe, and intervening if they go off course. Crucially, they always have to step in when the robots need to cross a road. This is time-consuming and limits how many robots can be out taking deliveries at any one time.
鈥淲e want to make them autonomous so anyone can order food from their app and the robot will deliver it within 30 minutes,鈥� says David Cardozo, one of Kiwi鈥檚 data scientists.
To solve this problem,聽since last month Kiwi has been using a , released in beta by Unity, one of the leading video game聽software companies. The system lets Kiwi鈥檚 data scientists create a simulation of a busy road that a computer Kiwi robot will try to cross.
As the robot repeatedly tries to cross the road, it uses machine learning to gradually teach itself the best techniques to get across unscathed. After a few thousand attempts the model was able to come up with a technique that helped it dodge between the onrushing traffic, shuffling back and forth in the road to keep itself safe.
The firm then conducted successful tests letting the real robots cross the road on their own by using this method聽and now plan to聽roll the software out across聽its fleet of 25 robots.
Early next year the service will arrive at Stanford University and eventually Kiwi wants the robots to be cheap enough聽to operate in cities across the US. 鈥淲e want to be able to escalate to hundreds or thousands in cities,鈥� says Cardozo.
