How Robots are Learning to Do Your Chores
CC by 2.0 BY CHRIS BARTLE
Machine Learning for Household Chores
How robots are learning to do your chores
In the not-too-distant near future, a robots might be folding your clothes, doing your dishes, or mowing the lawn. In fact, robots exist today with the mechanic capability to perform all these tasks — it’s just that there are an incredible amount of variables in tasks as simple as picking up a glass and filling with water. Humans don’t have to think about all these little decisions, but robots have to be told exactly what to do in order to complete a task. But these bots first have to actually learn how to do all your tedious household chores before we can all start living like the Jetsons.
One option is to program robots with how to react to thousands and thousands of scenarios, but there’s always the possibility that they’ll run into a situation they weren’t pre-programmed for. There’s a more efficient answer: machine learning. Instead of programing robots how to perform individual tasks, researchers are working on ways to teach robots how to think and adapt to different scenarios. It’s the 21st century version of the old “give a man a fish” versus “teach a man to fish” parable: give a robot a task, and it’ll perform it until it runs into the limits of its programming. But teach a robot how to think, and the possibilities are nearly endless. Machine learning is being used to bring us all closer to a future without chores. So how exactly are these robots learning how to do all this?
Through trial and error: Robots are programmed to try different tasks until they find a way that works — and then learn from their mistakes.
Elon Musk and Sam Altman began OpenAI, a startup that’s dedicated to producing AI software for free. Instead of actually building the bots, they’re hoping to create the brains — and to do that, they’re making algorithms that learn from extreme trial and error, known as reinforcement training. One project is to teach robots from Fetch Robotics how to perform common household tasks using this method. The hope is that when these robots are faced with new types of chore, they won’t require reprogramming. Embodied Intelligence, staffed with former researches from OpenAI, hopes to do something similar and develop the complex algorithms that let existing robots learn tasks on their own. These bots could learn to install car parts that aren’t like the parts they have installed in the past, or sort through a bucket of random holiday gifts.
The days where you can relax on the couch with a cup of coffee as your robot tidies up after a Friday night celebration aren’t quite here yet, but they’re closer than you might think, if these researchers have anything to say about it.
In virtual reality: Robots learn to complete household tasks in a computer simulation before trying it in real life.
The trial and error method is sometimes cost- and time-prohibitive, so other researchers are trying to use virtual realities to teach robots how to complete chores and tasks before letting them try it in real life. This has the benefit of letting humans make tweaks in real time to programming without damaging any real-life toasters or washing machines.
MIT and the University of Toronto have released a paper demonstrating how one system called VirtualHome could be used to train robots to execute household chores. The researchers turned verbal descriptions of household tasks into codes, which were then combined into programs representing more complex actions and then fed into VirtualHome where virtual robots executed them. In the future, these codes can be pooled in a database for robots to draw from. “Instead of each task programmed by the manufacturer, the robot can learn tasks just by listening to or watching the specific person it accompanies. This allows the robot to do tasks in a personalized way, or even some day invoke an emotional connection as a result of this personalized learning process.” said Qiao Wang, a research assistant in arts, media, and engineering at Arizona State University.
Another virtual simulation teaching robots how to perform tasks is AI2-THOR, an open-source virtual training ground. THOR (The House Of inteRactions) enables AI agents to learn how to interact with objects in familiar home settings such as kitchens and bedrooms without the possibility of damage to real-life property.
From each other: Someday, bots will be able to teach each other to do our chores.
Instead of learning from their own trial and error, or from human programming, another exciting possibility for machine learning is for robots to learn from each other.
In 2015, a robot in Cornell learned how to pick up cups and arrange them. Researchers uploaded the information into a cloud, where a different robot — with different physical parts for gripping and lifting — learned from the first how to pick up cups, and successfully arranged them in a different environment. This project is part of an effort to figure out the best ways for robots to share info with each other. This reduces the need for detailed programming, and might even let robots someday use knowledge from each other to react in unfamiliar situations.
To kickstart this, researchers have started with a goal both simple and ambitious: for robots across the globe to learn how to pick up one million different objects, upload the knowledge for how to do so into the cloud, and let other robots use this data. “We have powerful algorithms now—such as deep learning—that can learn from large data sets, but these algorithms require data,” says researcher Stefanie Tellex, an assistant professor at Brown University. “Robot practice is a way to acquire the data that a robot needs for learning to robustly manipulate objects.”
Whether robots learn from trial and error, through virtual reality, or from each other, a future involving machine learning is imminent. Today’s research is bringing us closer to the day when the burden of household tasks will fall to robots, and the chores of today will be a memory.