Researchers at MIT say a computer algorithm that will help robots recognize objects in a household environment could bring us one step close to robot "maids" to take the drudgery out of housework.
Such robots will need to be able to identify the objects they're supposed to handle and manipulate, the researchers note, and the mobility of such robots can help them look at those objects from more than one perspective to aid in that identification.
However, learning to meld those different views of an object presents a computational challenge, one addressed by an algorithm that can identify four times as many objects as algorithms using a single perspective, with make fewer misidentifications, the researchers say.
Developing that algorithm into a new version has yielded one that is up to 10 times as fast, which could help household robots which must make real-time identifications of target objects, they report in the Journal of Robotics Research.
Lawson Wong, a graduate student of computer science and electrical engineering, and his fellow researchers started with an algorithm created for tracking systems such as radar, which need to be able to determine if objects imaged at different instants are indeed the same object.
"It's been around for decades," Wong says. "And there's a good reason for that, which is that it really works well. It's the first thing that most people think of."
That algorithm generates numbers of hypotheses of which object in any one image corresponds to objects in other images, but problems arise as the amount of those hypotheses increases rapidly when images from new perspectives are added.
The algorithm copes by throwing out all but the most likely hypotheses at each new image step, but even still sorting through all candidate hypotheses at the end of the process is time-consuming.
The researchers at MIT went for a different approach; their improved algorithm doesn't discard any of the hypotheses, but doesn't attempt to examine every single one, either, instead opting to sample them at random.
In a sufficient number of samples, the researchers say, hypotheses showing significant overlap will be adequate to yield an accurate correspondence between object in any two consecutive images.
While the algorithm is a significant step forward, we're still a ways from robots that could set our dinner table with plates, glasses and silverware and then clean up afterward, Lawson acknowledges.
"As it is now, it's still very far from commercialization," he says.
Still, it suggests some day we may be able to leave the dinner table and just say, "let's let Roberta the Robot tidy up, shall we?"