Researchers at MIT have developed a new navigation method for robots that could be very useful for companies working on autonomous last-mile deliveries. In short, the team has figured out how a robot can figure out the location of a front door without being provided with a specific map beforehand.
Most autonomous last-mile delivery robots today essentially meet their customers at the curb, including the “cooler on wheels” developed by Starship and now adopted by a number of other companies, including Postmates. Mapping isn’t the only hurdle that will keep future delivery robots from driving to the door, just like the humans who make those deliveries today.
MIT News points out that it would be incredibly difficult to map an entire neighborhood with the accuracy required for true doorstep delivery—especially on a national (not to mention global) scale. Since that seems unlikely, and especially unlikely that every company looking to build autonomous delivery networks would have to handle deliveries separately, they set out to develop a navigation method that would allow a robot to process cues from its environment on the fly to figure out the location of a doorstep.
This is a variation of what you may have heard called SLAM (Simultaneous Localization and Mapping). The MIT team’s innovative twist on this approach is that instead of using a semantic map for the robot to identify and label objects in its environment, they developed a “cost-to-go” map that uses data from training maps to color-code the environment in a heat map. This allows the robot to determine which parts are more likely to be near a “front door” and which are not. Based on this information, it can instantly plan the most efficient route to the door.
It’s a much, much simplified version of what we do when we encounter new environments that we’ve never seen directly before: you know just by sight what is likely to be the front door of a house you’ve never seen before, and you know that essentially by comparing it to your memory of previous houses and the layout of those properties, even if you do all of that without even thinking about it.
The deployment is just one use case for this type of intelligent mapping of the local environment, but a good one that may actually be used commercially sooner or later.
