HAL, meet NIO.space
On the wall of our office are some hand-drawn maps of the lunar surface made for NASA in advance of the Apollo landings. It has always bemused me that 60 years later in the same unremarkable, easily overlooked yellow-brick building on the other side of the tracks from London’s MI6, we are using AI for the same task, but for Artemis. This building is ‘Arthur C Clarke’ House, purchased by science fiction rockstar Arthur Clarke, for the venerable British Interplanetary Society; our friendly landlords for over a decade.
Arthur C Clarke House could also be said to be the home of HAL (Heuristically Programmed Algorithmic Computer). While not the first cinematic space AI, HAL is certainly the most famous and the most prophetic foretelling of the emergence of this promethean technology in space. (As an aside, when we first started working with NASA thirteen years ago, a senior physicist playfully referenced HAL and announced, “AI in space is a really bad idea - have you not seen how 2001 ends?”.)
If you’ve marvelled at the Chess scene in 2001: A Space Odyssey - where astronaut Frank Poole abdicates erroneously to the algorithmic certainty of HAL (a phenomenon known as automation bias) you will know that movie explores many of the contemporary ethics of placing responsibility into the virtual hands of machines that give a perfect illusion of conscious thought, but don’t actually think.
Perhaps for this reason, we’ve generally never really got around to talking about our achievements in AI in space. However, over the past two years or so, we’ve been pushing the SOTA with numerous world (and space!) firsts, including the first ML trained in space, the first segmentation from orbit, the first migration of ML between instruments in orbit, the first federated learning between orbit and ground, the first virtual instruments and first unsupervised change detection from space - so it isn’t a stretch to say that the ‘real HAL’ is starting life in HAL’s shadow at Arthur C Clarke House.
So is AI in space a force for good? Should we be making the connection between HAL and putting ML onboard spacecraft?
At the end of last year, FDL.ai / Trillium and Oxford researcher Researcher Vít Růžička shared some of our ML onboard capabilities at the first TED conference on AI TEDAI as well as results from his PhD. Vít’s TED Talk is now live and we’re super proud to have these results being showcased on the world stage.
We call these capabilities, NIO.space.
HAL, meet NIO
When naming, we also went down the three letter acronym route and we call our AI for space NIO (Network Intelligence in Orbit). You might ask what’s the significance of bringing up HAL apart from a quirk of circumstance?
One of HAL’s failings was that it was a single point of failure. HAL was certain of its infallible logic; made with incomplete and contradictory inputs. One dataset, one model, one imperfect outcome delivered with ultracrepidarian confidence.
placeholder - WIP
The vision of NIO.space is to always use multiple datasets, multiple models; orchestrated, compared and continuously validated to create a coherent hybrid observation from multiple vantage points. This is the foundational architecture we’ve been testing to build NIO. While not perfect (and no AI should claim to be) building this kind of architecture from the outset is far less likely to be prone to HAL-like errors. If different instruments and different models are getting the same result, then we can feel more confident about the outcome. If one instrument and model makes a mistake, others can follow up to refute or validate an outlier.
STARCOP
A concrete application of the NIO approach is shared in Vít’s TED AI talk. STARCOP is a pipeline that automatically detects methane plumes from orbit using machine learning on multispectral and hyperspectral data to create a multi-instrument observation capability. One model says “hey, check this emission” and another instrument and model follows up and says, “I’ve taken a closer look with my better sensor and yeah - that’s a methane plume”.
This work is now being implemented at the UN Methane Alert and Response System (MARS).
Vít and the STARCOP team proved it was possible to identify excessive ‘super emitters’ from hyperspectral satellites using ML for the first time, supporting action to reduce greenhouse gas (GHG) emissions and achieve net-zero targets by allowing satellites to work together in concert. The findings have been recognised as one of Nature’s Journal Top 100 papers of 2023.
Vít’s PhD is also paving the way for a revolution in the way ML is used to extract insight from hyperspectral instruments in orbit. Previously, extracting insight from hyperspectral data was a hand-crafted effort requiring a preprocessing step and hefty compute requirements which made it untenable for onboard processing. This new architecture can extract insight from space derived hyperspectral data in seconds, paving the way for a revolution in the way spacecraft can work in concert to build a picture of our changing world, something we are looking to deploy as we improve NIO.space.
Reflecting on the legacy of HAL - the chess scene, the Pod Bay doors - and many others, we might agree that AI as a single point of failure in space operations is a bad idea. For this reason, we are building a networked, continually learning intelligence for space applications that is secure, respectful of data privacy, validated constantly and focused on helping us tackle the multitude of 21st century curve-balls coming down the pipe.
We believe that networked, orchestrated AI in space is a good thing, if we design it to be constantly checking its homework, staying humble, cross-validated and continually improving - moreover that the humans working alongside NIO are completely aware that even if an AI sounds confident, it’s always prudent to get a second or third opinion.
This is a lesson that applies on Earth as well - and one we should all be teaching our kids.
For that, we can thank HAL.
This research was supported by ESA and we’re grateful for the partnership of the Phi-Lab, D-Orbit, Unibap, NASA and Oxford University.
Did you know that FDL.ai and Trillium have over 50 other applied AI firsts? You can read about them here: https://trillium.tech/a-decade-of-ai
Vít’s research link: https://arxiv.org/abs/2410.17248
Vít’s TED Talk titled is ‘How AI helps us track methane — from space’ and you can watch here.
Disclaimer: While this article suggests that our offices at the British Interplanetary Society serve as a spiritual home for HAL, we acknowledge that HAL 9000 is a copyrighted character from 2001: A Space Odyssey. The name, likeness, and specific characteristics of HAL 9000 are the intellectual property of Warner Bros. Entertainment Inc. Any references to HAL 9000 in this article are made for discussion and commentary purposes only and do not imply ownership, affiliation, or endorsement by Warner Bros. Entertainment Inc. or the estate of Arthur C. Clarke.