Artificial intelligence (AI) is quite a trendy topic these days, especially since Google Alphago’s victory over the world champion Lee Sedol has given a bright illustration of the potential of machine-learning. Today, everyone seems to focus on the conversational branch of AI (bots, like the ones we build at SAP Conversational AI, or chatbots), but tons of other applications remain mostly unknown. This is why I’ve decided to dedicate this paper to a subject I’m passionate about: space exploration.
Indeed, it appears that AI can be an extraordinary boost to the discovery of our universe, for instance when it comes to navigation systems, situation analysis or even data transmission. So let’s try to anticipate some of the big chances that lie ahead! I’ve listed some of the most obvious, but if you want to go deeper into the matter, you can refer to Daniela Girimonte and Dario Izzo’s great paper on Artificial Intelligence for Space Applications or browse the NASA’s Jet Propulsion Laboratory webpage to check out the latest releases.
How can AI help us take a new leap in space exploration?
Transmitting data and analysing large quantities of it
Actually, space exploration already benefits hugely from AI. For instance, have you ever wondered how data is transmitted from one planet to another?
Until a few years ago, it used to be managed with a scheduling software operated by humans, but there were so many constantly changing variables such as the orientation of the spacecraft, the space-ground communication bandwidth, … that losing data, sometimes forever, happened all the time. Now, AI has been doing the job since 2005: the MEXAR2 system created by ESA is able to determine which data packets can be lost in case of memory conflicts, so that you never find yourself missing the most important piece of the puzzle.
There’s another key aspect when it comes to dealing with data collected in space: the sheer volume. A satellite like ESA’s ENVISAT produces 400 terabytes of data per year. If you want to process such a vast quantity of data the usual way, you might not even be able to use it before the next space exploration… this is why scientists have created a network of computers, each one of them receiving a small pack of data and processing it with AI, before regrouping all the pieces together.
Think 400 terabytes a year is a lot? Wait till the Square Kilometre Array telescope is working: it will be producing 720 terabytes a day! If we don’t invest a lot in AI programs capable of processing such amounts of data, the telescope will just be useless!
Saving time by giving more autonomy
One of the main difficulties when it comes to spacefaring is that – at least for now – many operations are still driven from earth. As most of you have probably seen Matt Damon’s Martian, let’s take the example of the red planet.
Depending on the relative positions of the planets, Earth is between 6,5 and 44 light-minutes away from Mars; I’ll let you imagine how long it can take to transmit one single message and wait for the instructions to come back. Of course, engineering teams are doing a great job anticipating every possibility and its appropriate response, but what if something unexpected happens? You can’t just rely on a message to be sent to Earth and come back with the answer you desperately and urgently need.
Consequently, the space industry will tend to look for more autonomy in the devices they use, which is where AI can be a powerful tool. For instance, we often use AI to detect and learn from patterns: in this case, you could apply these techniques to the analysis of sandstorms on Mars, in order to predict their evolution and adapt directly, without waiting for a command coming from Earth.
The NASA has already begun to work on that path: for instance, thanks to an AI software called AEGIS, the Mars Rover can select targets on its own for its laser and telescopic camera. This feature is “particularly useful at times when getting the science team in the loop is difficult or impossible — in the middle of a long drive, perhaps, or when the schedules of Earth, Mars and spacecraft activities lead to delays in sharing information between the planets” according to Tara Estlin, the leader of AEGIS development at NASA’s Jet Propulsion Laboratory. The NASA also equipped the Mars Rover Curiosity with autonomous navigation systems.
ESA’s recent International Rosetta Mission gave a good example on how space exploration could benefit from more adaptable systems: the mission almost failed due to a lack of adaptability.
In short, a device called Philae was sent to land on the comet 67P/Churyumov–Gerasimenko (nicknamed as “Tchouri”) in order to conduct a study. As you can imagine, landing on a fast-moving comet is not an easy task, and in that case it almost ended very badly: unfortunately, the comet’s gravity was slightly different from what it was supposed to be according to the preparatory calculations (with all the irregularities of the comet, a proper estimation of its gravity is particularly difficult).
The consequences were enormous: Philae hit the surface of Tchouri too strongly, bounced twice before eventually touching the ground (as a comparison, gravity on Tchouri is about one ten-thousandth of that on Earth). At the end of the day, on top on the damages caused by the shock, Philae found it had landed very uncomfortably on a 30% slope. As a result, some of the samples it was supposed to collect could not be harvested, and it was stuck in a part of the comet where it did not have enough sunlight to operate: only 1h30 every 12 hours. This reduced Philae’s autonomy, and also its ability to communicate with its bigger sister Rosetta, orbiting 200 kms above Tchouri.
We can very well imagine how AI could plug into space missions like Rosetta: by introducing a part of adaptability to the calculations and processes defined on Earth, based on what the spacecraft could directly observe. These things may seem obvious to you, but you need to know that most of the currently flying satellites are built on technologies more than 20 years old…
I’ve listed a few examples of the ongoing efforts in this field:
- The European Space Agency has obtained results to control a spacecraft around a small celestial body whose gravity field is unknown, using a technique called “neural reinforcement”.
- In order to foster the evolution of embedded systems, the NASA currently sponsors an application framework called ASPEN and based on AI techniques that can shelter different planning and scheduling applications, through a set of reusable software components that implement the elements commonly found in complex planning/scheduling systems, including: an expressive modeling language, a resource management system, a temporal reasoning system, and a graphical interface.
- The NASA also manages the Autonomous Sciencecraft Experiment (ASE), which has been operating onboard the Earth Observing-1 mission since 2003:
- onboard science algorithms are used to analyze the image data to detect trigger conditions (science events, interesting features, changes relative to previous observations, and cloud detection for onboard image editing);
- execution management software use the Spacecraft Command Language (SCL) package to enable event-driven processing and low-level autonomy;
- Continuous Activity Scheduling Planning Execution and Replanning (CASPER) software replan activities, including downlink, based on science observations in the previous orbit cycles.
Conclusion: the age of adaptability
Thanks to AI, the time when an organization would dedicate millions – even sometimes billions – of dollars for a system that couldn’t evolve during its use is over. NASA already relies on unmanned spacecrafts and devices to explore the farthest space. Even Elon Musk, who wants to inhabit Mars, has defined AI as probably the “biggest threat” to humanity, but will have to resort to it in order to fulfill his dreams. So whether you like it or not, until we find another and better solution, you’ll have to make do with AI for a while…
Rémi MEUNIER — SAP Conversational AI
I’d like to thank the Space Generation Advisory Council Team for their help, along with Sourav Karmakar, for suggesting me lectures and papers without which I couldn’t have written this article.