Project takes automotive AI chips into space
Hungarian automotive AI specialist AImotive is working on a prototype hardware platform to add artificial intelligence (AI) to space satellites. The chips already developed for automotive designs will work well with low cost, high volume satellites in Low Earth Orbit, says Tony King-Smith from the company.
The partnership with local space technology company C3S will see the chips in commercial space applications by the end of next year for telecommunications, Earth and space observation, autonomous satellite operation and docking support. Satellite operators such as OneWeb and SpaceX are planning to launch tens of thousands of such satellites over the next decade and companies such as Planet already run satellite imaging services for smart agriculture.
AImotive has been developing neural networks on automotive-grade hardware platforms for over four years and the same machine learning algorithms are well suited to the image recognition and decision making, says King-Smith. The designs are constrained by the power budget, memory size and communications bandwidth, and AI can help with all of these constraints by making decisions locally, he says.
“The work we are doing at the moment is studying the application to understand where the technology can best be used and what sort of characteristics of an AI engine make sense,” he said.
Xilinx last month launched an FPGA fto add AI to space system designs, but this is aimed at larger, geostationary satellites: FPGA PUTS AI IN ORBIT
The automotive qualification is suitable for LEO satellite designs, he says. “We are now looking for a bigger temperature range -40°C to +105°C and that’s helpful in any satellite with formal safety and this is all leading to take existing automotive chips and putting them into space,” he said. “For example the Apache 5 chip from NextChip in Korea is nicely on target for the end of the year and you could use that in a satellite.”
Next: radiation tolerance
Radiation tolerance is important, but less of an issue for LEO satellites that are designed to operate for three to five years. A key measure is the Single Event Upset (SEU), says King-Smith.
“SEU is something we do look at, that’s another thing we have been studying and that’s where the C3S expertise comes in, combining our knowledge of resilience and error correction to provide the required reliability,” he said. “We have been looking at SEU for some time in our automotive designs. You can get a [radiation] transient, it’s the same as a rogue bit from an image sensor.”
“We’ve also talked about the cost of developing a chip on a rad hard process, and then we can look at the pros and cons,” he added
”AImotive’s vast experience in deploying AI in highly constrained embedded environments is why C3S chose to adapt its aiWare NN acceleration platform for use in space”, said Gyula Horváth, CEO of C3S. “Adapting automotive or military grade tools and technologies for the space industry is not uncommon. The strong experience C3S holds in working with both cubic and large-scale satellites is an enormous advantage for this process and we are excited to be joining forces with AIMotive to make sure our space electronics platform will be the most robust on the market.”
C3S specializes in highly reliable 3-16U nanosatellite platforms supporting In Orbit Demonstration (IOD), Earth observation, space weather forecast, and IoT missions. In addition, C3S provides a turnkey solution for nanosat missions covering the whole lifecycle from mission planning to deorbiting. It also supplies the Electric Power Supply for the PLATO 2.0 medium-size satellite project.