What is the ultimate test of artificial intelligence in space exploration?

Imagine that an interstellar probe can choose its own orbit, take its own photos, and then send the probe to the surface of a distant planet without human help. This is an example of NASA's desire to use artificial intelligence, which has already used AI technology in missions on Earth and Mars.

It is understood that in December last year, NASA officially announced the discovery of Kepler 9 system, a planetary system consisting of eight planets around a star. Scientists discovered a galaxy with eight planets similar to the solar system for the first time in history. This discovery is inseparable from the assistance of AI, and NASA uses Google's machine learning to analyze kepler mission's data, achieving higher analysis efficiency.

In fact, NASA has already done a lot of things with AI.

Is artificial intelligence suitable for astronomical and cosmic work?

the answer is of course.

First, due to the high manned cost and high risk factor, such work as space exploration and planetary exploration is leaning towards unmanned;

Second, a large part of NASA's work is to analyze the image data sent back by sensors, and analyzing images is of course the best job of artificial intelligence.

thirdly, the fields of aviation and astronomy are highly digitized and informationized, which are suitable for mining historical data and training various algorithmic models to help scientists work.

Compared with the past, the biggest difficulty in using robots to explore the solar system and find signs of life is that they can't make intuitive and even creative decisions as effectively as humans.

The latest progress in artificial intelligence is expected to shorten this gap-there are no plans to send humans to explore the underground caves of Mars or look for hydrothermal vents in the icy waters of Europa, Europa. In their lifetime, these roles are likely to be replaced by smarter detectors and submarines. Even if they lose contact with the Earth for weeks or even months, they can withstand harsh environments and conduct important scientific experiments.

In the mid-199s, when Steve Chien took over the artificial intelligence team of NASA's Jet Propulsion Laboratory (JPL), the artificial intelligence at that time was more like science fiction, and no one could have imagined that it would play an important role in NASA's mission to mars in p>2. Chien has always had a wish to make artificial intelligence technology an indispensable part of NASA. However, at that time, artificial intelligence did not get enough attention. With less complicated algorithms running on old computers, technology was simply not competent for space missions.

however, Chien is very patient. His team is using technology to automate space missions and improve the work that has long relied on hard observation by researchers. For example, using the decision model of decision tree, JPL created the sky image classification and analysis tool (SKICAT), and used it to help NASA automatically classify the objects found in the second Paloma Mountain Survey in the early 198s. As long as SKICAT gets enough images for training, it can classify thousands of fuzzy and low-resolution objects in the investigation.

It is learned that after years of gradual improvement, when NASA asked them to design software for EO-1 satellite automation, Chien and his team made a breakthrough. In 23, NASA applied JPL's independent science and technology experiment (ASE) software to the satellite, and helped to study floods, volcanic eruptions and other natural phenomena for more than ten years. Before EO-1 was stopped in March, ASE software sometimes received alarms from other satellites or ground sensors, and automatically prompted EO-1 to capture images before people on the ground realized the incident.

JPL's work on ASE and other projects has given NASA confidence that artificial intelligence can play an important role in the Mars 22 mission. Chien and his team are developing a new type of rover, which is much more advanced than any other vehicle and can travel on the rugged surface of the planet. When looking for signs of life on Mars, the Mars 22 probe has considerable freedom in choosing research and experimental targets.

Recently, Chien, the head of the technical team of NASA's Jet Propulsion Laboratory and a senior research scientist in the mission planning and execution department of the laboratory, talked about the demand of artificial intelligence system for space travel in an interview with Scientific American. With the farther and farther vision of human exploration and the greater demand for intelligence, what will the "ultimate" artificial intelligence space mission look like?

The following is the finishing content

Q: Is the ASE software for controlling EO-1 satellite a milestone in NASA's AI application?

This is definitely a milestone of artificial intelligence, not only for JPL and NASA, but also for the whole AI ecosystem. That's because of ASE's great success and its longevity. This software is quite incredible-it has controlled the spacecraft for more than 12 years. During this time, it issued about 3 million instructions and made more than 6, observations, which actually achieved higher reliability than human operation of spacecraft. Such a success can actually democratize space resources. We have a web page where organizations all over the world can submit requests and send them directly to the spacecraft.

q: how many tasks is NASA willing to deliver to artificial intelligence?

one of the challenges that artificial intelligence faces at NASA is that it takes a lot of time and a long time to think about it because we are dealing with space missions. We must ensure that artificial intelligence is always in good working condition, that is, collecting scientific knowledge and protecting spacecraft. But that doesn't mean you can accurately predict what it will do. Some people want to get rid of this micro-management level and hope that artificial intelligence will become an assistant to scientists rather than a machine, because machines must be micro-managed. Some people are worried about replacing excellent scientists, but this is far from enough, so we don't have to worry about it.

q: how do you prepare to use AI to understand the unknown world?

unsupervised learning is very important for analyzing the unknown. A large part of what humans can do is to explain unfamiliar data. There will be many such problems at NASA. You will see some data, and some parts of these data are not suitable. Take Lewis and Clark's exploration of the northwest as an example. They didn't draw a map every 1 feet (which is what most detectors do at present), but Lewis and Clark's expedition described mountains, rivers and other features-putting them in the environment. We want artificial intelligence systems to do the same thing.

to develop such a system, we asked a student to take pictures with a digital camera during a cross-country flight. Then, we apply different unsupervised learning methods to our captured data. We want artificial intelligence to know that there are mountains, forests, rivers, and learning clouds, day, night and so on.

Q: What role does artificial intelligence play in the upcoming Mars 22 rover mission?

this task is applied to three aspects of artificial intelligence technology. The first is the autonomous driving technology of the rover, which can be traced back to the Pathfinder and is also part of the MER (Mars Exploration Rover) program. Autonomous driving is like a dial. You can strictly control it and tell the "rover" where to go, or you can let them drive. There are different trade-offs in terms of speed and safety.

the second area of artificial intelligence includes systems that will help rovers to carry out scientific research. The positioning ability will be much better, and there will be more instruments-not just the rover's SuperCam-which will provide imaging, chemical composition analysis and mineralogy. SuperCam is an evolution of ChemCam unique to the early Mars probe. The chemical composition of rocks can be understood by laser scanning and studying the gas produced. Previous Mars probes, Mars Science Laboratory and M22 now have more and more abilities to select targets and conduct subsequent image research according to scientific standards (such as the shape, texture or texture of targets). This ability is called autonomous exploration and collection augmented science (AEGIS) system, which enables the rover to conduct more scientific research in a shorter time.

Third, the "Mars 22" rover will also have a more complex scheduling system to make them more energetic. If the work is ahead of schedule or behind schedule, the rover will automatically adjust its journey, thus improving productivity.

Q: How does AI help the rover explore the caves on Mars?

When we explore the surface of Mars, scientists want to investigate lava caves on Mars. Because going deep into the cave is similar to a "relay race", such a task may only last for a few days, because the "rover" is completely powered by batteries, and cave exploration will require a lot of artificial intelligence. Artificial intelligence must coordinate, draw and explore as many caves as possible in a limited time. One of the methods we have been studying is dynamic area assignment, which may start in this way: you have four probes and want to walk 1 feet in a cave on Mars. The map of the first rover is to 25 feet, the second is 25 to 5 feet, and so on. They will gradually map the caves. This is a classic divide-and-conquer method.

they also use each other to transfer data from caves. Sending "rovers" into caves means they can't continue to communicate with the outside world. So they began to do what we call "Sneaker Netting"-the first "rover" entered the cave until it left the communication range and finished its observation, and then returned to the range to send data. The second car goes deep into the cave, but it only needs to return to the range of the first rover. In order to cover 1 feet, each detector gradually goes deep into the cave. Detectors don't come out of caves, but the data they collect can come out. This will be a three-day or four-day task, because the battery can only last so long.

q: what is the ultimate test of artificial intelligence in space exploration?

the ultimate test of artificial intelligence in space is "time". For example, the Europa submersible has to survive alone for several years, and may only come into contact with the earth once every 3 days. When you want to wait until the ice sheet melts before landing the submersible on the surface of the earth, it takes a year. In addition, the probe wants to find hydrothermal vents between the equator and the poles, just like a "rover" in a cave. In order to communicate with the outside world, it must go out and come back. In this case, the rover may live alone for six months or a year. To simulate this, we designed an AI-controlled submersible to study hydrothermal vents under the ice. Scientists want to study the effects of climate change under the Antarctic ice shelf-these tasks require similar technology.

Even so, this is nothing compared with the interstellar mission, because the spacecraft will run completely autonomously, and the communication journey to and from Proxima Centauri (the nearest galaxy) may last for nine years. If you go to Trappist-1 (a red dwarf star with extremely low surface temperature), this galaxy is the planet most likely to have life outside the solar system, about 4 light years away from us. Due to the delay of communication, the spacecraft depends more on itself, so when performing such a task, you need a strong enough AI.