Researchers Turn to Machine Learning to Help Robot Swarms Work Together
Caltech designed a data-driven method to control the swarms in cluttered and unmapped spaces without colliding with objects and each other.
Researchers at Caltech have designed a new machine learning system that enables robot swarms to work together through cluttered and unmapped spaces without colliding with objects or each other. Considering it's complicated enough trying to get drones to move within the same vicinity with others, the breakthrough appears to be a great accomplishment. Multi-drone coordination is a difficult process, which requires the drones to make split-second decisions concerning trajectories, most often without complete information on their future paths. Combine that with a cluster of drones in a confined area, and it's a recipe for disaster.
To overcome those challenges, the team developed a multi-robot motion-planning algorithm dubbed GLAS (Global-to-Local Safe Autonomy Synthesis), which imitates a complete information-planner using only local information. Think of it as painting a picture by extrapolating what the final image will be based on a fragment of a photo. The engineers then coupled GLAS with Neural-Swarm, a swarm-tracking controller augmented to learn complex aerodynamic interactions in a close environment.
"Our work shows some promising results to overcome the safety, robustness, and scalability issues of conventional black-box artificial intelligence (AI) approaches for swarm motion planning with GLAS and close-proximity control for multiple drones using Neural-Swarm," stated Soon-Jo Chung, Bren professor of aerospace at Caltech.
When GLAS and Neural-Swarm are used in tandem, drones don't require a complete picture of the environment it's traveling through, or even the trajectory others in the swarm will take. Instead, each drone learns to navigate any given space on the fly, and incorporate new information as they move into a new area. This allows for decentralized computing, letting each drone to "think" for themselves.
To test the new system, the researchers created a swarm 16 drones and flew them in an open-air arena, and found their machine learning algorithm could outperform similar platforms by 20% in most cases. At the same time, the Neural-Swarm outperformed commercial controllers, which are not equipped to handle advanced aerodynamic interactions and tracking errors.