The next phase of data collection: Expanding Acubed test flights across the U.S.
Airbus is pioneering efforts to develop autonomous flight technologies that will maintain Airbus’ exceptional safety records in the wake of a dramatic increase in air travel demand, predicted to double within the next decade. Today Acubed, Airbus’ innovation center in Silicon Valley, is announcing that it is expanding its flight test and data collection efforts to a significant number of airports throughout the United States.
Expanding test flights across the country Acubed announced its ambition to contribute to the development of safe, scalable, and certifiable autonomous flight systems for commercial aircraft some four years ago. Since then, incredible progress has been made on both sides of the Atlantic: in the last two years alone, we have validated our computer vision systems for use on Airbus’ Dragonfly demonstrator, developed our own unique simulation environments for AI model testing at scale, and helped demonstrate autonomous taxiing, takeoff and landing (ATTOL) with our partners at Airbus UpNext. These milestones have shown how the building blocks of autonomous flight promise to augment the abilities of human pilots, improve efficiency and increase flight safety. Now it is time for us to put these building blocks together.
To date, our data collection and testing efforts have primarily focused on the west coast of the United States. On sunny days, California blue skies present us with the ideal conditions for operating aircraft—on foggy days, the data we collect looks more like an ocean of gray. Data collected under all of these conditions has been helpful over the last year to build a solid foundation for autonomy, but our models have been limited to the specific geometry and surroundings of this one location.
The next step in our journey is to expand beyond the familiar surroundings and begin to roll out our data collection efforts to numerous airports throughout the country over the coming months. This will allow us to take the proven platform and process we have for collecting data, training models and testing the overall system and apply to new environments and conditions. Collecting data in these new settings will add to the large body of data we have already collected and hopefully reveal new ways that we can improve our models for the future.
Our foundational work in Silicon Valley We often talk about how the work we do at Acubed is like working at a startup. Much like any of the innumerable tech company startups in Silicon Valley, getting the “alpha” version right was vitally important. We needed to prove that our approach could work, and that we have a robust process in place before attempting to develop and test it at scale. For us, that has meant building solutions that can use a variety of machine learning and AI models to identify, predict and react to the environment around any future aircraft. These models will need to handle tasks ranging from interpreting visual and instrument data to flexible decision-making in a dynamic world. Robust machine learning and AI models, in turn, need lots of training data.
Data-driven AI for autonomous flight systems One of the biggest shifts in AI has been the move towards what many leading researchers characterize as “data-centric AI,” a move away from focusing on abstract model attributes and towards using real-world data as the primary method of improving model performance. For the last several years, we have been engaging in extensive testing with our data collection and ingestion processes to leverage this approach. Using a specially-modified civilian aircraft, we have been collecting flight data under a variety of flight conditions near our Silicon Valley innovation hub. This process has allowed us to fine-tune our ability to quickly collect the data we need, format the data for our model training pipeline and incorporate it into our model training process. Making this process efficient is vital, as the lessons learned from many of these first iteration cycles helped us make changes to the underlying software and hardware itself. These changes in turn made our data collection process go more smoothly, often making the next cycle of testing and iteration shorter, etc.
A key attribute of any future autonomous flight system is that it is able to react quickly and safely to unforeseen events, and in unpredictable conditions. So our initial training data incorporated flights during a wide range of weather conditions, altitudes, approaches, times of day, etc. These hours of flight time serve as a way of showing our decision-making models how the intelligence of human pilots translates to aircraft flight behavior under variable conditions that are recorded in the sensor data we collect. After establishing this body of data, we are then able to enrich the visual data, for instance, using synthetic data generation. We use synthetic data in an additive manner to simulate visual anomalies that interact with the physical properties of the aircraft’s sensor suite (e.g., lens flare), a wider range of unusual environmental light conditions (e.g., a solar eclipse), extreme weather anomalies (imagine snow in San Francisco…), examples of airborne and runway obstacles, etc. Populating our dataset with such extreme examples allows us to train our decision-making algorithms and then pressure test them within a simulated environment. Based on our outcomes over time, we can also evaluate and perfect our development process.
This approach to rapid iteration has allowed us to maximize the amount of testing we do early with the most flexible and lowest cost-of-development platform, before we one day expand to a large test aircraft.
We’re at an incredible inflection point in this industry as air traffic grows tremendously. Following current trends, we could be transporting 8.2 billion passengers by 2037 - the equivalent of today’s global population to keep up with the passenger demand and maintain (or increase) the present level of safety we enjoy, the industry will need more advanced technologies. Our vision is that autonomous systems will one day make air travel as ubiquitous as taking one’s daily trip up and down an elevator, but with the same sense of connection and wonder that air travel provides the world today.