Have you ever been curious about how insects manage to stray far from their homes and still navigate back? This question is significant not just in biology but also in developing AI for small, autonomous robots. Researchers from TU Delft drew inspiration from ants’ ability to visually identify their surroundings and their method of counting steps, ultimately creating a navigation system that mimics these behaviors for tiny robots. This approach enables these lightweight robots to return home after covering substantial distances while using minimal computational power and memory (only 0.65 kiloByte for every 100 meters). In the future, we may see these tiny autonomous robots used for various tasks, including inventory monitoring in warehouses and detecting gas leaks in industrial settings. The research team shared their findings in Science Robotics on July 17, 2024.
Advocating for Small Robots
Small robots weighing between tens to a few hundred grams possess potential for fascinating real-world applications. Their lightweight design ensures they are safe to use, even if they accidentally bump into people. Their small size also allows them to maneuver through tight spaces. If produced economically, they can be deployed in large numbers, enabling them to efficiently cover vast areas, like greenhouses, to quickly identify pests or diseases.
However, enabling these small robots to operate autonomously poses challenges due to their limited resources compared to larger counterparts. A significant hurdle is the need for self-navigation. While these robots can utilize external infrastructures, like GPS satellites outdoors and wireless communication beacons indoors, relying on such systems is often impractical. GPS doesn’t work inside buildings and can be inaccurate in congested urban areas. Additionally, installing and maintaining beacons in indoor settings can be prohibitively expensive or impossible, particularly in scenarios like search-and-rescue missions.
The AI required for autonomous navigation is often designed with larger robots in mind, such as self-driving vehicles. Some methods depend on heavy, energy-consuming sensors like LiDAR, which small robots cannot carry or power. Others use vision as a more efficient sensor that provides extensive environmental data. However, these methods typically seek to create detailed 3D maps, which demand considerable processing power and memory that exceeds what small robots can handle.
Step Counting and Visual Cues
This is where researchers have started looking to nature for solutions. Insects are particularly intriguing as they can traverse distances applicable to many real-world uses, relying on minimal sensing and computing resources. Studies have revealed much about the strategies used by insects, particularly how they combine motion tracking (known as “odometry”) with visual behaviors based on their low-resolution, nearly omnidirectional vision (referred to as “view memory”). While the concept of odometry is well understood, the specifics of view memory are still being explored. One early theory suggests that insects take “snapshots” of their surroundings. Upon approaching a familiar location, the insect can compare its current view with the snapshot and navigate to minimize differences, allowing it to return home while compensating for any odometry drift.
“Snapshot-based navigation is akin to the fairy tale of Hansel and Gretel,” explains Tom van Dijk, the study’s lead author. “Hans threw stones to avoid getting lost, but when he dropped bread crumbs consumed by birds, he and Gretel lost their way. In our case, the stones represent those snapshots.” He further adds that for the snapshot method to work, the robot must approach the snapshot closely. If the visuals differ significantly from the snapshot, the robot might end up lost. Therefore, it’s crucial to use enough snapshots—like Hansel needing to drop a sufficient number of stones. But, placing stones too close together would quickly deplete them. Similarly, for robots, using too many snapshots would require excessive memory. Previous research often spaced snapshots too closely, allowing the robot to navigate from one to the next.
“The key realization behind our approach is that snapshots can be set much farther apart if the robot uses odometry to reach between them,” notes Guido de Croon, a professor specializing in bio-inspired drones and co-author of the article. “Homing will be effective as long as the robot arrives near enough to the snapshot location, meaning that the drift from odometry needs to stay within the snapshot’s catchment area. This method enables the robot to travel much further since it moves slower when navigating back to a snapshot compared to traveling straight to the next one.”
The insect-inspired navigation strategy successfully allowed a 56-gram “CrazyFlie” drone, equipped with an omnidirectional camera, to traverse distances up to 100 meters using only 0.65 kiloByte. All visual processing was carried out on a compact device known as a “micro-controller,” commonly found in inexpensive electronic products.
Applying Robotic Technology
“This insect-inspired navigation strategy significantly advances the applicability of small autonomous robots in real-world scenarios,” concludes Guido de Croon. “While the strategy is not as robust as cutting-edge navigation techniques and only allows return to the start point without generating a map, this could suffice for many uses. For example, in warehouses for inventory tracking or in greenhouses for crop monitoring, drones can fly out, collect data, and return to their base. They could save relevant images on a small SD card for later analysis by a server, but wouldn’t require those images for navigation.”
