Researchers at the Delft University of Technology have developed the first-ever swarm of tiny drones that are capable of autonomously detecting and localizing gas leaks in tight indoor environments. To find a gas leak in a building or industrial site, human firefighters risk their lives since it can take a long time to detect the source. These new drones could have major implications in this area.
Designing AI for the Drones
The biggest hurdle for the researchers was to design the artificial intelligence (AI) needed for the complex task. Because of the tiny size of the drones, the computational and memory parts needed to fit tightly into them. The researchers relied on bio-inspired navigation and search strategies.
The research was published on the ArXiv article server, and it will be presented at the IROS robotics conference later in the year.
What’s Required for Autonomous Gas Source Localization
The task of autonomous gas source localization is extremely complex, and it requires artificial gas sensors which are not very capable in detecting small amounts of gas. They also struggle to stay sensitive to quick changes in gas concentrations.
Besides the actual task, the environment also causes problems when it is complex. For these reasons, traditional research has evolved around single robots that search for a gas source in small, obstacle-free environments.
Guido de Croon is Full Professor at the Micro Air Vehicle laboratory of TU Delft.
“We are convinced that swarms of tiny drones are a promising avenue for autonomous gas source localization,” says Guido de Croon. “The drones’ tiny size makes them very safe to any humans and property still in the building, while their flying capability will allow them to eventually search for the source in three dimensions. Moreover, their small size allows them to fly in narrow indoor areas. Finally, having a swarm of these drones allows them to localize a gas source quicker, while escaping local maxima of gas concentration in order to find the true source.”
Despite the benefits of these properties, they also make it difficult for engineers to implement AI into the drones for autonomous gas source localization. Because of the limitations around onboard sensing and processing, the AI algorithms used in self-driving vehicles are not applicable. Because they operate in swarms, drones also need to avoid colliding with each other while collaborating.
Bart Duisterhof performed the research at TU Delft.
“Actually, in nature there are ample examples of successful navigation and odor source localization within strict resource constraints,” says Duisterhof. “Just think of how fruitflies with their tiny brains of ~100,000 neurons infallibly locate the bananas in your kitchen in the summer. They do this by elegantly combining simple behaviors such as flying upwind or orthogonally to the wind depending on whether they sense the odor. Although we could not directly copy these behaviors due to the absence of airflow sensors on our robots, we have instilled our robots with similarly simple behaviors to tackle the task.”
The tiny drones rely on a new “bug” algorithm called “Sniffy Bug,” which enables the drones to spread out before they detect any gas. This allows them to cover large environments and avoid obstacles or each other.
Once one of the drones senses gas, it communcationates that to the others, which will then collaborate with each other to find the gas source as quickly as possible. More specifically, the drones perform a search for maximal gas concentration with an algorithm called “particle swarm optimization,” or PSO, where each drone acts as a particle.
The algorithm was inspired by the social behavior and motion of bird flocks, with each drone moving based on its own perceived highest gas concentration location, the swarm’s highest location, and its current moving direction and inertia. One of the benefits of PSO is that it only requires measuring the gas concentration without the gas concentration gradient or wind direction.
“This research shows that swarms of tiny drones can perform very complex tasks.,” says Guido, “We hope that this work forms an inspiration for other robotics researchers to rethink the type of AI that is necessary for autonomous flight.”