A team from the University of Zurich (UZH) trained the AI ââto fly a drone through a virtual environment full of obstacles before unleashing it in the real world, where it was able to sneak up to 40 km / h / 25 mph, three times faster than the previous best driving software. Principal researcher Davide Scaramuzza, director of the Robotics and perception group, explains that the work, carried out in partnership with Intel
An article describing the project, Learning to Fly at High Speed ââin Nature, is published this month in the journal Science Robotics.
“Our approach is a springboard towards the development of autonomous systems capable of navigating at high speed in new environments with only on-board detection and computation,” the document concludes.
Scaramuzza’s group previously demonstrated the first autonomous drone capable of beating human racing pilots on a course using sophisticated trajectory planning using external sensors. The new work takes a somewhat different approach and is stand-alone: ââmachine learning AI effectively has learning with an omniscient master in the virtual world and learns to imitate the technique of the master. (A bit unlike the Sparring in the Matrix programâ¦ except the student is an AI.)
In the virtual world, the learning AI piloting a drone is alongside an expert system that possesses “inside information”, detailed knowledge of obstacles and the exact position and speed of the drone, information that would not be known in the virtual world. the real world. This allows the expert system to plot the optimal route around obstacles. The machine learning system generates a large number of possible routes and as the flight progresses it learns which one is best from the master, updating its neural network each time.
After a large number of virtual flights, the learning AI was then ready to take on the real world. It flew in difficult situations where there were branches and other obstacles, or narrow openings to negotiate, with poor visibility and limited computing power.
The UZH pilot objected to two state-of-the-art drone piloting programs. At low speeds performance was similar, but at higher speeds other systems suffered from too much ‘latency’ – they were not able to calculate a safe route quickly enough – and suffered several accidents. The previous software used separate modules for detection, mapping, and planning, while the machine learning approach merges the three functions and provides faster routing.
The latency of the other two systems was 65 and 19 milliseconds for a sample task, the UZH system handled 10 milliseconds. While this may not be consistent for every driving challenge, it shows the potential of the approach.
Scaramuzza notes that this is the first time that such a system has been trained in a completely simulated environment and then used its learning in the physical world. While this can be useful for training delivery drones and air taxis to negotiate the cityscape in all weather and lightning conditions – inexpensively and without the risk of an accident – there may also be potential much more important.
The technique suggests that guided machine learning in virtual environments could be a way to quickly and inexpensively train all sports of robotic systems. For example, a robotic system for unloading boxes from trucks could get the equivalent of years of experience learning how to handle inconvenient shapes and sizes of packages. Home cleaning robots, which sometimes struggle with odd-shaped parts or particular furniture, could get smarter through virtual training.
In the immediate term, this approach is likely to be applied primarily to small drones, allowing them to perform aggressive maneuvers faster than any human pilot. The paper indicates that even faster speeds could be achieved with better sensors and more precise modeling; it is not clear what the absolute limits could be.
You can watch a two minute video showing how âLearning High Speed ââFlight in the Wildâ was done here.