The dream of autonomous flight will be game-changing and is arguably the key factor in making urban air mobility (UAM) vehicles accessible to a commercial market. After all, if each UAM required a pilot - whether on the ground or in the aircraft - we would be severely limited by availability.
This is where machine learning and AI come into play. They promise to be the backbone of the eVTOL sector and will help to reduce reliance on human involvement. But how will machine learning translate to the aerospace industry, and what difficulties does this present?
Machine learning is a form of artificial intelligence in which a computer analyses its behaviour to improve its processes. It functions much in the same way humans learn from mistakes and successes but in a computing context.
The process of machine learning relies on data and algorithms. You begin by feeding the computer training data, which defines the specific success you require. The computer passes this data through an algorithm, defined by the programmer. The type of algorithm depends on the data set, but could be one of the following:
Neural networks will likely be the preferred method of machine learning in an eVTOL context because they are “deep learning” algorithms that can process, understand, and learn from a large amount of data in many different layers. Understandably, autonomous flight will involve vast data sets related to environment, weather conditions, performance, and so on.
We already have many different versions of machine learning around us, ranging from simple to complex. Everything from our phones and apps to robot vacuum cleaners and spam filters uses some kind of machine learning. Considering we have gone from IBM’s Deep Blue to autonomous cars in less than three decades, the future of machine learning is only going to get more complex.
Machine learning is already in place in self-driving cars, where it serves to render the environment and predict possible changes that might impact its journey. Its process breaks down into four processes:
The vehicle captures data from images and LIDAR pings, which it passes through a series of layered algorithms. It uses different layers to ensure it does not miss objects; for example, if pictures are unclear it will detect objects from LIDAR signals. The vehicle can then pass the information through its decision matrix to determine what it should do.
For the most part, machine learning in eVTOLs will function in much the same way. They will use neural networks to plan routes, detect hazards, and deal with weather conditions that might impact their journey. However, the aerospace context presents some unique challenges not faced by ground-based vehicles.
Along with the “basic” feature of navigation, machine learning presents opportunities for aircraft optimisation. Machine learning’s potential in an eVTOL context is essentially unlimited, providing there is data to create a training program.
The aircraft will rely on a navigation system based on multiple cameras and LIDAR. As we discussed in a recent post, LIDAR will form the basis of this system, and machine learning will not only process the images but also help the system to learn from new hazards. Importantly, this will help to counteract the unpredictable nature of hazards in the air, such as birds and other eVTOLs. While it may never be able to completely predict the flight pattern of birds, machine learning will help to build awareness of this kind of hazard.
Another key area that machine learning will impact is battery optimisation. In our post about UAM battery technology, we discussed the current limitations the technology presents: lifespan, charging time, and so on. Providing it receives the right data, a neural network will be able to understand how best to use the battery, when and how it can reduce usage, and the best time for charging.
A similar system is currently in development by Bosch and DiDi Chuxing for driverless cars. They have created a “battery in the cloud” service that gathers data from the vehicle to understand and optimise battery performance. Smart software analyses battery performance and provides the driver with tips on how to extend its life. This kind of software can be adapted to remove the human element and simply have the computer learn from its own advice.
Machine learning will also help to create effective emergency strategies. Swiss-based AI company Daedalean AG is currently working on exactly these systems. They are currently teaching their neural network using offline data gathered from real flights and are specifically avoiding the typical “learn on the job” model of machine learning.
They want their systems to not adapt themselves mid-flight, as this could become dangerous. Instead, they are filling their systems with data gathered from manned flights to assess various risks and develop strategies before the aircraft ever takes flight. So far, they have found this model outperforms human pilots every time.
The early stages of eVTOLs will likely keep a link to a ground-based network for their information processing. Currently, driverless cars link to the cloud to stream and manage data - some send up to 25GB of data every hour. It will be the same for autonomous aircraft, although as processing power and speed increase, we may see all data manipulation take place in the aircraft.
In the eVTOL context, machine learning will do a lot of work. Not only will it plan and run the intended route, but it will also learn from previous flights to improve operation and hardware performance. Of course, its ability is still fairly limited and its improvement presents a range of unique challenges.
Perhaps the most obvious challenge machine learning - and the wider use of AI - faces is regulation. Luuk van Dijk, co-founder of Daedalean, considers machine learning to be a mystery to most, and this could impact its rollout in a commercial setting. This will make it incredibly challenging to pass legislation that regulates machine learning because the parameters of what it covers will constantly change as machines update themselves.
Next is the issue of public acceptance. UAMs present unique challenges for acceptance because few people will feel confident putting their safety in the hands of an unmanned aircraft, particularly one that has complete control over itself. However, safety will be a key factor in regulation, and machine learning systems will likely have various redundancies to make them as safe as possible.
AI will also have to overcome challenges presented by the weather. Normal aircraft are vulnerable to weather changes, and this issue is compounded when the aircraft are smaller, lighter, and flying around cities where winds can reach massive speeds.
An answer to this issue may already exist in the aerospace market. In early 2020, Delta Airlines announced its use of machine learning to create and predict weather patterns along its key routes. Their system will allow them to analyse, locate, and avoid turbulence and other occurrences that lead to flight cancellations.
This kind of system could theoretically find use in an eVTOL context. Machine learning could locate key areas of turbulence - particularly around tall buildings - and at the very least avoid them. The same process could overcome things like snow, hail, etc., as the aircraft could adapt energy usage, rotor speed, and other factors that impact how it tackles difficult weather.
Machine learning in eVTOLs is still in its development stage. However, the more we learn from its application in other contexts, the easier it will be to apply it to unmanned aircraft. Much of the machine learning technology used in driverless cars applies to eVTOLs, providing it is adapted to an aerospace context. While regulation will play a large part in its future development and application, it will be a case of laws catching up with technology rather than them setting the standards.
