|Landing drones may look easy, but it’s really not. It could encounter complex turbulence or irregular motion of air generated by the airflow from each of the drone’s rotor / Photo by: alexsalcedo1 via 123RF|
Landing drones may look easy, but it’s really not. It could encounter complex turbulence or irregular motion of air generated by the airflow from each of the drone’s rotor as the ground grows closer during the descent. This turbulence could only be some annoying bumps or it could be severe enough to throw the drone momentarily out of control, causing structural damage.
Drone Landing and Takeoff: Why Are They Tricky
Such complex turbulence is not yet well understood and is also a challenge to reduce even for the autonomous mini flying bot. This is the reason why landing and takeoff are often two of the trickiest parts when it comes to drone flight. Usually, drones inch slowly and wobble toward the landing ground until the power will be finally cut off, dropping the last distance left to the landing spot. But a new system from the California Institute of Technology (Caltech) enabled drones to land more quickly and smoothly.
|Such complex turbulence is not yet well understood and is also a challenge to reduce even for the autonomous mini flying bot / Photo by: feverpitched via 123RF|
The Use of Deep Neural Network
A team of artificial intelligence experts and control engineers from the California Institute of Technology collaborated to design drones that can fly smoothly even when it is near the ground and can overcome the complex turbulence. In a report published by Science Daily, the team used what they call as the deep neural network so drones can “learn” to land quickly and safely while using less power than normal drones did in the past. A deep neural network (DNN) is a network that comprises the output layer, input layer, and another hidden layer in between the two. Each of these layers does certain ordering and sorting to deal with unstructured or unlabeled data. This is why it has the word “deep” in its name because it means data inputs are being churned in various layers. Furthermore, DNNs are ideal to be used in repetitive tasks because they can perform automatic learning techniques.
|A team of artificial intelligence experts and control engineers from Caltech collaborated to design drones that can fly smoothly even when it is near the ground and can overcome the complex turbulence / Photo by: unitysphere via 123RF|
The Caltech team dubbed their system as the Neural Lander. They said that it is a “learning-based controller” that can determine the speed and position of the flying drone and then function to edit its rotor speed and landing trajectory accordingly in a way that it will achieve the smoothest landing possible.
Caltech Jet Propulsion Laboratory Division of Engineering and Applied Science's Ben Professor of Aerospace Soon-Jo Chung, whose research focuses on aerospace robotics, space autonomous systems, and distributed spacecraft systems, explains that their recent project can be used in drones so they can fly more safely and smoothly even in the existence of unforeseeable wind gusts.
The Spectral Normalization Technique
Soon-Jo Chung and co-lead authors said that to ensure that drones can fly smoothly under the DNN guidance, they used the spectral normalization technique. This technique smooths the output of the neural net so it will not make differing predictions, as conditions and input shift. At the same time, the team has measured the drone landing improvements by studying the irregularity from that of an idealized flight. They first tested it in the 3D space.
Testing the Drone’s AI System
A total of three tests were conducted by the Caltech team. These were the (1) straight vertical landing, wherein the drone lands vertically, (2) the descending arc landing, and (3) when drone quickly glides across the edge side of the table, wherein the effect of complex turbulence from the surface ground varies sharply. These tests were conducted at the three-story aerodrome of the Center for Autonomous Systems and Technologies, Caltech’s research center meant to create autonomous systems and advance fields of bio-inspired systems, autonomous exploration, and drone system.
The result of their tests showed that the AI system was able to decrease up to 100 percent vertical error. This paved the way for controlled drone landings. In the same way, the AI system reduced 90% lateral drift. Their experiments moreover showed that with the use of an artificial intelligence system, drones performed actual landing instead of stopping first at 10 to 15 centimeters above the landing ground prior to the drop.
Caltech’s Assistant Professor of Computing and Mathematical Sciences Yisong Yue also emphasized that Neural Lander was able to achieve smoother and speedier landing and only committed less error than traditional drones during their experiments. Moreover, it was able to glide smoothly above the ground.
The Potential Application of the AI
A patent for the said AI system has already been filed by the Caltech team. They foresee that aside from commercial applications, the system can also be crucial in the field of medicine and healthcare. For example, drones can be equipped with AI that can smoothly and swiftly land in locations that are not easy for humans to locate. It can also be used for medical transport. One of the research head for a separate “air ambulance project” and present CAST’s Hans W. Liepmann Professor of Aeronautics and Bio-inspired Engineering Morteza Gharib likewise commented that the significance of landing smoothly and swiftly cannot really be overstated if the need is for transporting injured people.