Artificial intelligence in zero gravity

Southeastern USA at night. Photo courtesy of NASA, ISS030-E-55569.

As I consider the potential for artificial intelligence and machine learning to transform existing industries, few opportunities loom larger on the horizon than space.

Satellites orbit the Earth to collect and transmit data and the industry hinges on the effective management of that data flow. From ingress (remote sensing) to egress (data downlink) to whatever processing happens along the way, how that data flow is managed is a significant differentiator between the legacy and new space companies.

Legacy players launched behemoth satellites with billion-dollar project budgets to collect a lot of data and downlink via one fat connection. These satellites downlinked terabytes of data that was analyzed and actioned, often long after the fact. They have essentially built a library of largely evergreen, dated data.

New players invert this paradigm, launching small satellites in low Earth orbit to collect and relay megabytes of data continuously, and analyze and action that data as close to real-time as possible. They are building something more akin to the Internet, tapping into the ephemeral data of everyday actions.

But getting to real-time action on satellite data remains a challenge. Once a satellite collects data, it downlinks that data to a ground station antenna on Earth, and then from the ground station the data is piped to a data center or directly to the end consumer where it can be processed. The physical distance that the data travels on Earth can vary drastically, from hundreds of feet to tens of thousands of miles. Once processed, an action is taken based on what the data reveals. Consequently, the response time or that lapse between data collection and data-driven action can be significant — hours, days, months, or even years.

We are experiencing an exponential explosion in the number of connected devices on Earth, and consequently, that bottleneck between collection and action appears to be tightening without any material signs of relief.

Not all data is created equal, however, and today all space data is transmitted as if it were just that, equal. Machine learning and other forms of artificial intelligence have the potential to relieve pressure on already overburdened infrastructure and fundamentally alter the way that we view and utilize space assets. It could drastically reduce response times and bring a transformational level of autonomy to systems that today require our undivided attention.

How might this impact the average person? Consider a service like OnStar, where you physically push a button in your car which contacts a Help Center if you are in distress. This is a manual intervention — push button (action), and a voice materializes in your car (reaction). This was an innovation built for cellular networks, and consequently, only works when cellular service is available. Imagine driving down a highway, tired and ready to be home. Just ahead of you, around a bend in the road, three cars were involved in a collision. Depending on your reaction time, this could mean your own collision, or at the very least, frustration from being caught in a traffic jam. The latter — frustration — is your best case scenario, and something like OnStar is of no help.

Now imagine this same scenario with intelligence built-in and connectivity available anywhere you travel. Satellites are the only technology that can provide true global connectivity. In this scenario, a notification automatically appears on your dash stating that a collision occurred, and assuming options are available, reroutes you automatically to your final destination. Best case, you end up at home without frustration, much less a scratch.

Satellites are ubiquitous. Add a further layer of intelligence and they will transform the way we engage with the physical world. This is viable today and does not involve crowdsourcing traffic data, looking at your phone while driving, or waiting till you hear sirens to understand why you are now in standstill traffic.

Here are a few advancements that have me particularly excited:

  • Algorithm Development. Developing for space means working with constraint — a small form factor, low-power, remote hardware device suffering from variable-latency and sustained radiation, among other challenges. One cannot simply equip satellites with the newest NVIDIA GPUs. The development of new algorithms optimized to run on existing low-memory, low-power embedded hardware is the path forward, and will become commonplace in a relatively short time.
  • Intelligent Filters. Software-defined radios (SDRs) are the workhorse of all small satellites, acting as the gatekeeper for both mission control and normal operations. Integrating a deep learning model into the SDR could, among other things, create an intelligent filter between the data collected and the data downlinked, effectively freeing operators from the paradigm of treating all data as equal.
  • AI Chips. GPU is in high demand, though its power requirements and energy output, as well as its form factor, hinder adoption in small satellites. Fortuitously, a new cohort of companies has emerged with transformative chip architectures meant to vastly increase processing performance while simultaneously diminishing both the power required and the energy output. These companies are offering increased performance at lower utility cost, and may ultimately birth an entirely new class of small satellite.

Each of these has the potential to be transformational, and ALL will influence satellite resource allocation which alone will have a real economic impact downstream (i.e. value for you and me).

Satellites are a unique class of hardware. Once launched, there is no servicing of the equipment — other than via software — which creates plenty of opportunity for entrepreneurs looking to build machine learning and artificial intelligence applications.

Machine learning can be applied to many areas of satellite operations to drive improvements in resource utilization and performance, including:

  • Autonomy. Managing a constellation of small satellites spread across a single orbital plane, let alone multiple orbital planes, is no small order. And the challenge is only exacerbated as a globally distributed network of satellites expands in concert with a globally distributed network of ground stations. One can throw only so many bodies at this massive orchestration challenge before looking to software. This is a ripe area for disruption as cross-talk between satellites (sat-to-sat), heterogeneous satellite networks and diverse ground stations becomes reality.
  • Communications. The communications stack leverages the interface between software-defined radios and ground stations. It is a complicated physics and orchestration challenge. Where do you locate ground stations? How accurate must the satellite pointing be to close the link with the ground? How often must you downlink to balance storage capacity on the satellite with mission parameters for data collection and frequency of available downlinking opportunities?
  • Power Management. The Electronic Power System controls the collection (solar panels), storage (rechargeable batteries), and metering of all power onboard, including in eclipse (i.e. when the Earth passes between the satellite and the Sun, cutting off new solar power generation). Informed power allocation can make or break a mission.
  • Satellite Health. Space is hard. Low Earth orbit is objectively less so, though the hardware is still put under constant strain. There is low-level sustained radiation slowly degrading systems (e.g. flipping bits), single upset events that can fry an entire satellite, and errant space objects that threaten collision at more than 4.53 miles per second. The command line interface is virtually your only window into what is happening onboard the satellite, and the unknown unknowns are plentiful.

AI and machine learning will further transform the space industry. The question that remains, however, is who will reap the greatest rewards from these advancements, and to what degree will satellite operators themselves effectively productize the new intelligence they unleash.

If you like what you read be sure to share on social or give the author a round of applause!