A promising future for AI and autonomy in space
What you will learn:
- Make spacecraft more autonomous.
- How Machine Learning and Deep Learning Drive Self-Reliance Algorithms.
- Leverage software design tools such as MATLAB and Simulink for algorithm training and other aspects.
Artificial intelligence (AI) continues to dominate the news, but reality hasn’t always lived up to the hype. This is the case in the automotive industry, where the goal of fully autonomous vehicles for the general public remains elusive.
Surprisingly, the space industry, long a mainstay of conservative engineering, stands ready in the shadows to deploy machine learning in a practical and pragmatic way and once again become a leader in computer applications. Indeed, AI could be a competitive differentiator for the next generations of spacecraft.
AI and its potential in the space industry
The applications of AI in the space industry lend themselves well to gradual development. First, a spacecraft that can perform the analogous task of “driving on the empty road” has already solved the problem of autonomous driving, to some extent. No vehicle, pedestrian or bicycle in space can jump in front of a spaceship.
Second, while the primary purpose of a car is transportation from point A to point B (which the AI focuses on activating), a spaceship often has many other purposes. In fact, the equivalent of the function of “transport” (called guidance, navigation and control in the parlance of the space industry), while critical, is often only the catalyst for other functions, such as provisioning. satellite images or the realization of scientific observations. This means that a multitude of functions can benefit from AI, some more risky than others. As AI has not yet proven itself in space, the less risky functions are good candidates for its first applications.
The space industry also has a long history of gradually making spacecraft more and more autonomous. Space vehicles that can make more decisions on their own are much more valuable, especially for space exploration.
For example, NASA’s Curiosity rover is armed with an instrument called ChemCam, which analyzes the composition of Martian rocks and soils. But in order to do this, ChemCam must first aim at a target. Giving pointing instructions from the ground is a tedious process, limited by the fact that the right communications satellites are within sight of Curiosity and even the time it takes for controls and data to travel from Mars to the United States. Earth (called the light time constraint). For this reason, Curiosity uses an autonomous targeting algorithm to aim its instrument during times when ground command is not available.
It represents a good example of a “low risk, high reward” application of AI in space. This approach also has very few drawbacks: the algorithm adds scientific value if it succeeds, and if it fails, little or nothing is lost. However, since the addition of autonomy also necessarily increases the complexity of the design, such algorithms are difficult and expensive to create using traditional programming techniques. But the expansion of the use of these algorithms will result in a better performing and more competitive industrial landscape.
Machine learning and deep learning advance autonomy
Recent advances in AI, driven by machine learning and deep learning, have made self-reliance algorithms not only more powerful, but also faster to create and accessible to more businesses. and engineers. These advancements also occurred at a time when the space industry is experiencing unprecedented growth in the number of new private companies entering the market.
Earth sensing is one of the rapidly growing space applications, with a demand for satellites capable of detecting the Earth’s surface using cameras, radars and even radio frequency analysis. This data is used for a variety of purposes, from agriculture (assessing crop health) to finance (counting cars in shopping mall parking lots to draw conclusions about the economy).
A common problem that Earth sensing satellites struggle with is that the amount of data they can collect is often more than the amount of data they can transmit to Earth. Additionally, some of the data collected is not really useful for its intended use (e.g. did it capture an image of the target, or were there clouds in the way?).
Sorting payload from “wasted” data is now done by ground-based computers, increasingly using deep learning techniques. It would be much more efficient if the payload data could be separated from the unnecessary data by the satellite itself, which would then be able to send only payload data – and more – to clients on Earth. Deployment of AI algorithms using deep learning on board the satellite enables this data separation. Companies able to do this will have an advantage in the market and engineers with these skills will be in great demand!
Given that there is a long history of increasing the autonomy of spacecraft from the earliest days of spaceflight, and that machine learning and deep learning can facilitate the resolution of some of the common problems of spaceflight. space industry, the industry is well positioned to successfully apply AI. Many argue that the industry’s inherent conservatism and risk aversion will prevent the adoption of AI.
There is a flip side to this argument, however. These same characteristics can help the industry determine the best ROIs for AI, without dramatically increasing the risk for expensive space vehicles and missions; for example, by carefully selecting the appropriate functions to apply AI techniques, as in the case of the Curiosity Rover. The culture of the industry is also changing, with new companies ready to try new things.
AI and engineering software
Once AI is adopted aboard spacecraft, its use could become more widely accepted. More critical use cases for deep learning include vision-based navigation for rendezvous and proximity operations as well as hazard detection and avoidance for lunar or planetary landings. In addition, engineering software will continue to play a key role in the advancement of AI.
Aside from machine learning and deep learning for a moment, even traditionally coded AI algorithms benefit from tools that drive the design of increasingly autonomous spacecraft. The trend is to allow spacecraft to possess complex decision-making capabilities that were previously left to a human operator. One of these examples is to interrupt or continue an autonomous docking between two spacecraft.
Modern engineering tools abstract from design by providing views that obscure complexity, allowing engineers to better understand potential system behaviors. For example, modeling decision-making logic as digital executable state diagrams rather than if-then-else statements in code allows examiners to gain a better understanding of possible logical paths. This design abstraction approach, provided by tools such as MATLAB and Simulink, helps engineers understand the impact of individual functions on system behaviors during design phases.
Despite improvements in design abstraction, design errors in these complex AI systems can be subtle and difficult to detect. For example, assessing the robustness of a vision-based detection and perception algorithm under conditions involving lighting and perspective is difficult by examination alone and requires extensive simulation and testing.
Design tools like MATLAB and Simulink are most effective when they provide capabilities not only to train algorithms more easily with features such as automated labeling, but also for early verification and validation by simulation. This allows requirements-based testing well before the algorithm is deployed to hardware.
Machine learning and deep learning, the next frontier of AI, inherently requires engineering software, as machine learning and deep learning models are generated by computers, not humans. . However, the software tools used by data scientists to prototype machine learning and deep learning techniques often have a steep learning curve and are not always designed for production deployment of algorithms, which may lead to inefficient development cycles.
Thus, there is a growing need for aerospace engineers to have access to tools such as MATLAB and Simulink for machine learning applications they are already familiar with that are designed for engineering workflows, as opposed to the data science.