Are Artificial Intelligence and Machine Learning Essential to the Success of the Commercial UAS Industry?

The use of artificial intelligence (AI) and machine learning (ML) is no longer siloed to military applications. As the commercial unmanned aircraft systems (UAS) industry advances, the use of AI and ML is becoming more common and some experts in the industry claim it will not be successful without it.

According to a report from the Brookings Institute, AI generally refers to machines that respond to simulation in ways similar to traditional responses from humans. AI creates these responses using algorithms that are created from data. ML looks for trends within AI data. These two technologies together can be used to enable intelligent decision making.

Tel Aviv-based software provider Airwayz has created an unmanned traffic management (UTM) system for UAS that utilizes AI. It is currently being tested in a pilot program in Israel where five drone companies are all flying their fleets within the same corridor, something that has never been trialed before. The UTM communicates with newly developed UAS service systems (USS) to manage the airspace, Tomer Serok, a director at Airwayz, told Avionics International.

The UTM has access to a lot more information than the USS like weather and non-flight zones, Serok said. Once the UTM receives a request from the USS, it begins sorting that information in order to approve or disapprove of the flight plan and it does this by using AI.

“Basically, the UTM replaces the software of the ATM, the air traffic management, the one you have in the control tower, but also the men in the control tower so the decision making is done by the UTM as well,” Serok said. “The USS is requesting a certain flight path from the UTM, the UTM is checking that comparing it with an offline zone, with weather conditions, with the capability of the drone itself because it could be multiple drones, and if there's any conflict with other USS.”

This is all done pre-flight. While humans can be in the loop at the USS level, Serok said they cannot at the UTM level because there is just not enough time to account for human decision timelines.

“Without artificial intelligence, this thing will not work,” Serok said. “The drones will never be able to deliver you burgers, pizzas, do an inspection, police work, firework, in a city. It’s not going to happen. If you just got to do it manually, which means you have to predefined corridors for drones for the operator, you might get to 10 percent or five percent of the capabilities, but you will never even get close to the potential of the drones. You cannot do drone deliveries because if you're doing everything very rigid there's no way you can squeeze all these things into the sky.”

Zipline, a drone logistics company, relies on AI to run its drone logistics network, Conor French, general counsel at Zipline told Avionics International through email. The network uses AI in route planning, flight control, flight safety decisions, sensor fusion, recovery systems, and automated air traffic control.

“We now fly the equivalent of the circumference of the Earth every day,” French said. “We use this flight data to refine and improve the performance and reliability of our technology. Our AI has 15 million km of real-world flight heritage, all with zero injuries, and over 1.5 billion km flown in simulation. Over time, we’ve increased our fleet size and airspace density significantly. For example, when we started operating in Rwanda in 2016, we safely flew two aircraft at the same time. We now regularly have 20 aircraft interacting with each other in close proximity in the airspace. We’ve improved delivery accuracy from a small field to an area the size of a few parking spots.”

Airwayz and Zipline are not alone in their use of this technology. SkyGrid and SparkCognition announced an AI-powered cybersecurity system for drones in January of this year. American Robotics also uses AI in its Scout System, which includes a base system and UAS, that received approval from the Federal Aviation Administration (FAA) for beyond visual line of sight operations (BVLOS) earlier this year.

“Unlike traditional anti-malware reliant on signatures of known threats, our AI models don’t require an existing threat database,” Ali Husain, SkyGrid’s chief software architect, told Avionics in January. “However, it’s still important to store data from a threat once it’s detected. This will allow our models to learn the DNA of each threat and detect similar threats that may emerge.”

While many companies in the industry are continuing to adopt this technology, there is still skepticism from some that AI/ML cannot be certified as safety critical.

“It was delightful to see a lot of misunderstanding and preconceptions about how impossible this [AI] was,” Luuk van Dijk, founder and CEO of Daedalean, said during a Revolution.Aero Town Hall last year. “…Our goal is to develop the kind of machine equivalent of human capability, which would call AI so that we can get to this ultimate form of autonomy."

Daedalean is creating avionics system piloted with AI for electric vertical take-off and landing (eVTOL) and general aviation aircraft. Daedalean and the European Union Aviation Safety Agency (EASA) published a report in April of 2020 which examines the use of neural networks in avionics.

“The trick is that worrying about the non-determinism is a bit of a red herring because the system itself is perfectly reproducible and analyzable,” van Dijk said. “What you have to do is you have to show that the data that you trained and tested on in the lab is sufficiently representative of the randomness you're going to find out in the real world.”

UAS companies like Airwayz and Zipline are gaining this all-important data through the tests and application of their products. Serok said Airwayz has completed 1,500 flight tests with its technology so far. French said Zipline’s has 15 million km of real-world flight and over 1.5 billion km in simulation flights.

This data is also helping inform the decisions of regulatory bodies where there is still much uncertainty.

“As novel technologies, or at least applications, there is complexity and uncertainty in the regulatory environment,” French said. “As a result, very few companies are operating at scale using these technologies and so regulators continue to face challenges of available data and information in determining how to appropriately assess risk. We are working to overcome these challenges by understanding where the various regulatory bodies are today and their vision for a successful UAS space and helping to provide as much information as we can about how our systems, products, and services work. This approach can vary by geography, but reflects the importance of education, awareness, and agreed-upon standards and mechanisms to oversight for the health and growth of the UAS space.”

In a recent webinar, Chip Downing, senior market development director at Real Time Innovation (RTI), suggested that a combination of the DO-178C avionics software standard, the Future Airborne Capability Environment (FACE) standards, and data distribution (DDS) standards could work together to certify these systems.

“They’re very tightly connected, and I really think especially on the vehicle itself, but possibly the management of the vehicles also, you're going to have machine learning capabilities,” Downing said.

Downing cited CoreAVI’s Vulkan certification that runs on top of general-purpose computing on graphics processing units or GPGPUs and the AI/ML engines on the GPGPUs.

“If you take a look, almost every major microprocessor manufacturer is now making an AI engine or some type of ML capability on these GPGPUs, which is a huge change from just a few years ago,” Downing said. “So now, not only can you use that information, you can now actually do safety certification evidence around that information using this Vulkan and CoreAVI technology, and I think that's a huge step to getting us to the point where you can actually start integrating AI into the systems.”

While companies are working with regulators to achieve regulatory acceptance, French said Zipline is also working to public acceptance of its technology.

“As people and communities come to appreciate (and potentially depend upon) the considerable societal benefits of these technologies there will be greater incentive for governments to support their broader usage,” French said. “This accounts for both the more traditional lens of enhanced safety made possible by integrating AI and machine learning in our systems and also how UAS can improve people’s access to healthcare, reduce environmental impacts of ground transportation and lessen the burden or volume of collisions on our roadways. This closely interrelated dynamic between regulatory approval/oversight and public acceptance has existed through the history of aviation and so it is likely to be a particularly critical interface in the UAS space in the weeks, months and years to come.”