Yesterday we announced Solo, our spacecraft intelligence model, and there are a few questions we’ve gotten repeatedly. So let's answer them! A Q&A 🧵

Q: "Why are you working on autonomy?"

A: We've spent four years building, testing, and flying deep space spacecraft. And we realized the traditional operating model is itself a huge constraint. Every single one of our deep space missions depends on access to expensive, scarce terrestrial dishes and extremely constrained data rates. That means constantly having to decide what information is important enough to send back to Earth, which is generally a fraction of the information that is available onboard. Solo moves more of that decision-making to the spacecraft, where the data already is, to increase mission reliability.

Q: "How can you possibly have enough data to train this?"

A: Here's the thing: we are not training a Waymo. The problem space is deliberately constrained; deep space is a much more predictable environment than a city street. Our subsystem models operate within narrow domains and we generate data constantly through testing and simulation. We also deliberately inject errors into our hardware-in-the-loop testing platform to train and test against edge cases that conventional testing can’t reproduce.

Q: "Why use AI for this at all? Why not deterministic flight software?"

A: We are using deterministic flight software! Solo does not replace any of the existing flight software, GNC algorithms, or other deterministic systems aboard the spacecraft. Our onboard deterministic algorithms already solve many problems very well.

Take trajectory planning. We already have a physics-based algorithm that can calculate a trajectory. There’s no reason to replace that with AI. But if the spacecraft completes a burn and discovers it isn’t where it expected to be, Solo can recognize that discrepancy, call the existing trajectory-planning algorithm to generate a new trajectory, and then determine what the spacecraft should do next.

Q: "Is Solo an LLM flying a spacecraft?"

A: No. Solo is not a chatbot and it isn't trained on language.

Q: "So is Solo a digital twin of the spacecraft?"

A: No. We use physics-based models and deterministic algorithms where they’re the right tool, but Solo itself isn’t a digital twin of the spacecraft.

Solo is an autonomy stack designed to reason across information from different spacecraft systems and determine what should happen next.

Q: "Can’t you just write rules for every failure mode?"

A: We already write deterministic rules for well-understood conditions, but you can’t anticipate, and write a rule, for every combination of conditions a spacecraft will encounter. There is a wide range of problems between conditions that are addressable with predetermined rules and problems no intelligence model could possibly solve. Experienced mission operators operate in that middle ground today, using the available data, along with their judgement and experience, to determine how to respond. Solo is an attempt to move this reasoning onboard.