If you’re visiting a friend and find yourself sleepily stumbling through their house in the middle of the night, basic common sense tells you that if you see a fridge, you’re probably in their kitchen. But a toilet? More than likely you’re in a bathroom. It’s a common sense approach to navigation that robots are finally able to copy to make them better at quickly finding their way around a home.
Researchers from Carnegie Mellon University working with Facebook’s AI research team have developed a new approach to an autonomous navigation system for robots known as SemExp (awkward, but less awkward than Goal-Oriented Semantic Exploration) that uses deep learning to not only train them to recognise objects but to also know where in a house they’re typically found.
Training a robot to find its way around your home isn’t impossible, but at the moment, it requires lots of manual input working in conjunction with image and object recognition so that the automaton knows what areas you’ve designated as a bedroom versus a kitchen, for example. But homes are ever-evolving, stuff gets left all over the place, and objects, such as a footstool that a robot was relying on as a recognisable navigation landmark, could be obscured by an abandoned pair of pants one day.
The goal of SemExp is to make robots more flexible by giving them a basic form of common sense. Machine learning allows them to not only recognise individual objects, such as a coffee table versus a kitchen table, or a dryer versus a dishwasher, but also where in a user’s home those objects are most likely to be found. Sure, some lucky homeowners might have a toilet conveniently installed in their office to maximum productivity. But for the most part, you’ll always find the big metal box called the fridge in the kitchen, or the oversized glowing panel on the wall in the family room.
One of the ways robot vacuums have dramatically improved their performance and run times is by intelligently navigating a home so that they’re not cleaning the same areas twice, or by heading directly to a room that hasn’t been recently tidied. This research has similar goals, and the researchers hope they can make robots better at navigating a home so that they’re not reliant on specific objects being in specific places all the time.
The improved adaptability also promises to make robots easier to interact with. If you ask a robot to fetch you a cold drink from the fridge, it will know that the kitchen is the most likely place it will find that appliance, and it will do its best to head straight there first. But even its autonomous route planning will be improved as a result of having a better understanding of objects around the house. If the robot spots a dining table, it could extrapolate that it’s currently near the dining room, which are typically adjacent to the kitchen, allowing it to plot a shorter route overall. Most importantly, this research could one day make robots the ultimate tool for finding where you left your keys, or who stole the TV remote.