A household robot can learn how to do almost any chore in about 20 minutes when taught by a human using an iPhone camera and a litter picker.
Robots tend to perform well only on specific tasks that they have been trained for, like sorting rubbish or picking up laundry, and can quickly run into difficulty when encountering a task in unfamiliar environments, such as in homes.
Advertisement
Now, at New York University and his colleagues have created a way to teach robots that involves a litter grabber equipped with an iPhone, a system called the Stick.
A person uses the Stick to demonstrate to the robot what they want done so it can learn to mimic it (see video, below). 聽鈥淵ou take out your iPhone from your pocket and put it on, and you have a tool that, for all intents and purposes, is practically the same as the robot,鈥� says Shafiullah.
The robot and the Stick system are collectively called Dobb-E 鈥� a reference to Dobby the free-house elf from the Harry Potter franchise, both 鈥渂ecause the system is doing household chores and also because we are trying to set it 鈥榝ree and open source'鈥�, he says.
Each new task takes about 5 minutes of instruction, or about 24 demonstrations, and it takes a further 15 minutes for the robot鈥檚 algorithms to learn it. Then the robot can perform the task by itself.
Shafiullah and his team tested Dobb-E in more than 20 New York apartments on 109 different tasks, such as opening cupboards and doors or taking a towel off a rail, and found that the robot had an average success rate of 81 per cent.
Tasks were relatively constrained to where Dobb-E learned them. Learning to pull a chair out from under a dining table, for instance, didn鈥檛 transfer to teaching the robot how to pull a chair out from under a desk. However, Shafiullah hopes that, as the robot collects more data and its learning algorithms are improved, it will be able to look at different tasks and generalise beyond specific problems.
鈥淭here is no internet for robot data, and that鈥檚 what we鈥檙e hoping to change,鈥� says Shafiullah. 鈥淥nce you have some cheap way of collecting data at scale, then you can see all this data and do a million different things with the robot, hopefully.鈥�
Although the demonstration stick is relatively cheap, at $25, the actual robot costs about $20,000, so it could be a little while before Dobb-E is seen in homes, says Shafiullah.
鈥淭he sheer diversity and open-world nature of the tasks that Mahi and team have demonstrated is impressive,鈥� says at the University of Washington in Seattle. 鈥淚 expect that these types of dataset collection techniques and fine-tuning methods will only get more popular with the advent of scaling in robot learning.鈥�
arXiv
