Intel and ESA launch first AI-enabled satellite, PhiSat-1 – Business Insider
This week the European Space Agency (ESA) and Intel announced they had successfully put the first satellite with on-board AI-processing into space, PhiSat-1.
The PhiSat-1 uses Intel’s Movidius Myriad 2 chip, which was not originally designed for space travel.
PhiSat-1’s on-board AI is able to select and automatically delete photos of Earth if they are too obscured by cloud.
AI has become mainstream on Earth, but getting it onto satellites has been a huge challenge, and PhiSat-1 could pave the way for innovations in how satellites detect natural disasters, or even how we communicate with Mars rovers.
On September 2, a satellite the size of a cereal box took off for space.
Names PhiSat-1, its mission was to monitor polar ice and soil moisture, making it — at least superficially — a fairly unglamorous piece of kit.
But for the satellite’s creators — the European Space Agency (ESA), chip giant Intel, and Irish robotics company Ubotica — this launch represented months of work, and had been postponed by a failed rocket launch, two natural disasters, and a global pandemic.
It was also a huge technological leap forward.
PhiSat-1 is the first satellite to make it into orbit with AI onboard — and the tech could transform the way we respond to disasters, such as oil spills and wildfires.
Why has it taken so long to get AI into space?
Gianluca Furano of the ESA told Business Insider that although AI was popular for building products on Earth, getting the tech onto satellites has been a massive challenge.
“To use AI in a data-critical application is not straightforward even on Earth,” he said. Flying an AI-powered chip like the one onboard PhiSat-1, 329 miles above the planet, means any repair, software patches, or upgrades will be that much more difficult.
“The complexity simply explodes,” Furano said.
On top of that, radiation in outer space makes it hard to build an AI chip that can withstand being in orbit.
“Whatever has silicon inside is disturbed by ionizing radiation,” Furano said. This means computer chips that would function perfectly on Earth would have unworkably high error rates, and could even catch fire if sent into outer space without shielding or modifications.
But while lots of satellites carry custom-built spaceworthy chips, PhiSat-1 uses Intel’s Myriad 2 chip, a commercially available chip found in a handful of commercial drones — and even the headset for Magic Leap’s AR goggles.
Ubotica took the Myriad 2, adjusted its software, and built some electronics capable of shutting it down if it looks likely to overheat. Researchers tested it by taking it to the largest particle accelerator on Earth at CERN in Switzerland.
By May 2019, the satellite was ready, but it took longer than planned to make it into space. The original launch was scheduled for September 2019, but the rocket meant to take PhiSat-1 failed to launch, prompting an enquiry.
While that was ongoing, the coronavirus pandemic set in — and to top it off, two hurricanes hit the launch complex in French Guiana and a ground station in South Korea.
On September 2, PhiSat-1 finally made it into space, and has been successfully orbiting ever since, putting its AI into practice.
PhiSat-1’s AI lets it ignore clouds — which is a lot more impressive than it sounds
Like the satellite itself, the function of PhiSat-1’s on-board AI doesn’t seem glamorous at first, as it essentially auto-deletes photos of Earth it deems too cloudy.
Before PhiSat-1 sends pictures back, its AI decides whether they are free enough of cloud cover. If the AI decides more than 70% of a photo is obscured by cloud, it automatically scraps that image.
This is useful because it saves Phisat-1 a lot of energy. Ubotica Chief Technology Officer Aubrey Dunne told Business Insider about 67% of the Earth’s surface is normally covered by clouds, so this means the AI is cutting a lot of images straight away, rather than using processing power sending useless images back to scientists.
“The cloud detection is about data reduction,” he said.
This could pave the way for fighting fires and driving Mars rovers
Furano said that while the successful launch of Phisat-1 with it’s cloud-scrapping AI is exciting, current applications of the tech are “low-hanging fruit.”
“The real exciting part comes when you think of applying onboard processing with machine learning [the type of AI PhiSat-1 uses] to applications where you need a very prompt answer,” he said.
As an example, Furano said AI could massively speed up satellite alerts of disasters like oil spills.
Ubotica’s Aubrey Dunne said the company is already working on adapting the technology to build satellites which could automatically detect wildfires and flares from oil refineries. “We’re trying to do latency reduction, so we’re trying to use AI to perform applications where the time to get an answer is really important, and fire detection is a great example,” he said.
“You want to try to alert authorities and relevant people on the ground in the relevant places about the location and extent of the fire, and how it’s changing, moving, and shifting without having to wait a day for the data to download and wait another day for it to be processed on the ground,” he said.
Dunne thinks AI could have off-world applications as well, for example speeding up communications with rover robots on Mars.
Ubotica is aiming to get the next Myriad AI chip into space in early 2022. “It’s a few satellites about twice the size of the PhiSat satellite, and it’s partly ESA-funded but also partially commercially funded, so it’s a nice sort of bridge to the commercial side as well,” Dunne said, although he did not elaborate on where the commercial funding would come from.
Jonathan Byrne, head of Intel’s Movidius technology office, told Business Insider that Intel sees AI satellite chips as a nascent market, but one that promises to grow.
“The technology’s 20 years old that’s on satellites, and 20 years ago GPUs were starting to grow massively — and you’re starting to see the same thing,” said Byrne.
Furano said the tech currently in orbit is 15 to 20 years behind the processing power that can be found in smartphones today, meaning getting new, improved chips is top of his list. “We desperately need AI in space,” he said.