Machine Learning Lets Swarms of Drones Fly Without Colliding

In the not so distant future, robots may very well be a major component of modern human life. Whether we look at the driverless revolution to the use of drones, we can expect to see these machines to inhabit the space around us fairly soon. However, with several robotic devices functioning simultaneously, a greater issue emerges- the risk of collision.

Multi-robot situations will become an increasingly common occurrence, and so engineers will need to ensure they can exist in the same space without hindering each other. In the scenario of drones, the machine needs to be capable of making instant decisions to adjust its trajectory to prevent an accident, whilst also making sure it doesn’t lose the target destination. These sorts of changes need to be made in seconds whilst also recognising that the situation may change as a result of a movement from the other drone. This process has a vast array of complex elements that need to all be accounted for to prevent a collision, and thus allow the drones to function appropriately.

Attempting to resolve this, a group of scientists from the California Institute of Technology have developed technology designed to aid robotic movement in close proximity with other robots. Their solution is two-pronged. First, there is Neural Swarm- a controller for multirotor swarms. It helps combat aerodynamic issues that arise when drones are in the same vicinity e.g. downwash of air from higher vehicles to lower one helping maintain suitable distances between the vehicles. Secondly, the Global-to-Local Safe Autonomy Synthesis(GLAS) algorithm plans suitable motion for the robots. The product? The system can vary the behaviour of every machine based on data collected about the local environment of each drone. The robots are therefore capable of making real-time adjustments to their flight based on data this data they are continuously interpreting and prevent drone collisions during close proximity flight.

GLAS at work

Caltech’s Bren Professor of Aerospace, Jet Propulsion Laboratory Research Scientist Soon-Jo Chung stated, "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."

The team put the entire cyber-physical system to the test by performing a test on 16 drones in a confined space. The results highlighted the effectiveness of the programme with GLAS performing 20% better in comparison to the previously best motion planning system.

Thumbnail Credit: Caltech

Video Credit: Caltech