AI enables drones to fly into the unknown
When it comes to exploring complex and unknown environments such as forests, buildings or caves, drones are hard to beat. They are fast, agile and small, transport payloads and use sensors to get virtually anywhere. But without a highly accurate map, autonomous drones can hardly find their way around an unknown environment. To fully exploit their potential, they currently still need experienced human pilots.
“When manoeuvring a drone, you need to understand the environment in a fraction of a second in order to quickly steer the drone onto collision-free paths,” says Prof Davide Scaramuzza, who heads the Robotics and Perception Group at the University of Zurich. “This is very difficult for both humans and machines. Experienced pilots can reach this level after years of training. But machines still struggle with it.”
In a recent study, Scaramuzza and his team trained an autonomous quadrocopter to fly at speeds of up to 40 km/h through previously unknown environments such as forests, buildings, ruins or trains without colliding with obstacles of any kind. The drone relies only on the built-in cameras and the calculations of its on-board computer.
AI requires a learning phase. The Zurich researchers implemented this by training the drone’s neural network: the missile learned to fly around obstacles by observing a kind of “simulated teacher”: an algorithm that flew a computerised drone through a simulated environment full of complex obstacles. The teacher algorithm was aware of the quadrotor’s position and its sensors’ readings at all times, and had enough computing power to calculate the best trajectory in a fraction of a second.
The data from this simulated teacher is used to teach the neural network how to predict the best trajectory based on the data transmitted by the sensors. This, the Zurich researchers say, is a big advantage over existing systems that first use sensor data to create a map of the environment and then plan trajectories within that map – two steps that take a lot of time and make it almost impossible to fly at high speed right away.
After training in the simulation, the system was used directly outdoors, where an autonomous drone was able to fly at speeds of up to 40 kilometres per hour in various environments without collisions. “This means that artificial intelligence can use high-performance simulators to achieve comparable navigation capabilities much faster than a human, virtually overnight,” says Antonio Loquercio, PhD student and co-author of the paper. These simulators do not have to be exact replicas of the real world. With the right approach, even simple simulations are sufficient.
The applications of the system are not limited to quadrocopters: The same approach could be useful to improve the performance of autonomous cars, for example, or even to train AI systems in areas where collecting data is difficult or impossible.
The next step is to improve the system and develop faster sensors that provide more environmental information in less time, so that the drone can fly safely at speeds above 40 kilometres per hour.