For drones trying to navigate a busy environment like a warehouse or a forest at high speed, the ability to know exactly where they are at all times would seem pretty essential. Not so, say researchers from MITs Computer Science and Artificial Intelligence Laboratory (CSAIL), who have a devised a new, efficient way to guide drones around obstacles. The key ingredient? Uncertainty. With most drones — and, indeed, most self-driving vehicles — navigation starts with a map. To draw one, depth sensors are used to scan the immediate environment which is compiled into a single 3D model. This then tells the vehicle not only where they are at any given moment, but also how to get to their destination. Its a method commonly known as simultaneous localization and mapping, or SLAM. SLAM has served the community pretty well to date, but it has its downsides. For one, its a very intensive process, that needs lots of high-fidelity data and computing power to process it. This is why Waymo and Ubers recently settled lawsuit was all about LIDAR — the laser-firing sensors used to collect and process depth data. Data is important. But, this process creates problems at high speeds and with small crafts like drones. They dont have the time to collect all the data they need, and giving them the processors to understand it all is expensive. But one way to bypass these requirements, says CSAILs Peter Florence, is to plan less and react more. Florence and his team have developed an alternate method of navigation and obstacle avoidance that is tuned to these demands, which they call NanoMap. It still works by collecting 3D data about the environment, but this information is never fused into a single map and is instead stored in a series of snapshots. This allows for faster reaction times: the drone is processing less information each second (thats the uncertainty aspect), and so crunches the data with ease. Because were not taking hundreds of measurements and fusing them together, its really fast, Florence tells The Verge. And when we want to plan our way around the world, we just search back through the views we already have. There are drawbacks to this method, too, and Florence says NanoMap isnt great for applications that need high-quality maps of their surroundings (think, drones doing surveying work in agriculture or helping with search-and-rescue missions). Similarly, the makers of self-driving cars will likely be happier with SLAM, given that its hardware demands are less of a burden in vehicles that are going to be big and expensive anyway. But, says Florence, for small drones, NanoMap could be the perfect way to give them obstacle-sensing abilities without overtaxing their digital brains. Early tests in both real-life environments and simulations are promising. We have drones that can fly around now, but theres nothing thats as good as a hawk, for example, that can just blaze through a forest at top speeds, says Florence. When youre going that fast, you just cant expect to have perfect geometric measurements of the world. You need something else. I think just working on robots that can move really fast and gracefully is fun, so thats the main thing.
Somehow, that helps it avoid hitting trees. The current generation of autonomous drone navigation and flightpath planning systems are almost too precise, demanding hundreds of measurements be taken so that the UAV knows exactly where it is in space at any given moment. And if those readings are off by even a little, then the drone is in for an impact. What's more, all that data collection is computationally intensive -- especially for smaller drones where the space and weight capacities are limited. The new NanoMap system from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL), however, strikes the right balance between accuracy and speed. With it, drones can navigate heavily populated areas -- think forests or Amazon fulfillment centers -- at up to 20 mph. Simply put, the system doesn't sweat the details. Unlike other common mapping systems, such as simultaneous localization and mapping (SLAM), which are data intensive and difficult to maintain at real-time, the NanoMap uses depth-sensing to measure just the drone's immediate surroundings. This enables the drone to understand generally where it is in relation to obstacles and anticipate how it will need to change course to avoid them. "The key difference to previous work is that the researchers created a map consisting of a set of images with their position uncertainty rather than just a set of images and their positions and orientation," says Sebastian Scherer, a systems scientist at Carnegie Mellon University's Robotics Institute, wrote in an MIT release. "Keeping track of the uncertainty has the advantage of allowing the use of previous images even if the robot doesn't know exactly where it is and allows in improved planning. " This uncertainty is surprisingly helpful. Without working the factor into its modeling, MIT's test drone would crash roughly 25 percent of the time whenever it drifted more than 5 percent away from where it expected to be. But by incorporating that uncertainty, the MIT team was able to reduce crashes to just 2 percent of its flights.