The Physical AI Revolution Will Start in Space

Modern AI has been built around a single powerful axiom: capability increases with scale. Frontier labs have embraced that principle with enthusiasm, training models on increasingly large datasets, with more parameters and larger context windows, all powered by enormous data centers. 

That scale has been possible in part because AI has largely operated in the digital world, where both the inputs and outputs – language, code, images and other forms of information – are themselves digital. 

The physical world presents a different problem. There is no internet-scale dataset representing every possible state of the entire physical world. Companies pursuing general-purpose robotics must contend with unfamiliar objects, textures, and shapes; dynamic environments; and a range of possible tasks. The number of conditions a model must understand, and the number of possible interactions between them, quickly creates a combinatorial explosion. Building intelligence that can operate across that world requires enormous amounts of data and compute.

And yet we believe the next wave of AI is physical: intelligence that can reason across information about the physical world and act on it through hardware. The question is how to make the problem tractable, and what the first instantiation of that wave will be. 

At AstroForge, we believe the answer lies in deep space. A spacecraft is a complex system, but it is fundamentally constrained: by hardware, by physics, by a defined set of actions, and by a deterministic logic that already governs much of its behavior. That’s the insight behind Solo, our spacecraft autonomy platform. Solo uses specialized models to reason across what is happening on the spacecraft, then use bounded, rules-based decision-making to turn that understanding into action. 

Millions of miles from Earth, the spacecraft needs to not only understand its state, but translate that into commands that the physical hardware can safely execute. That’s why we aren’t throwing out the deterministic logic spacecraft already use. But those rules aren’t enough. They need intelligence. 

Why spacecraft? 

Spacecraft are already very good at executing predefined logic, and engineers write deterministic rules for a range of conditions in advance. The problem comes from those combinations of conditions engineers can’t anticipate. Today, human operators bridge that gap. Mission operators examine telemetry, and use their experience and judgement to decide how the spacecraft should respond. Solo is our attempt to move that reasoning on board. 

General-purpose physical AI faces a much broader problem. Autonomous vehicles must reason about and interact with other vehicles, roads, and pedestrians; general purpose robots may need to perform a wide range of tasks across changing conditions; and autonomous drones must navigate weather and other obstacles in a crowded airspace. Generalization across those situations is tied to the value of the product. 

But a spacecraft is different, because we already understand the machine and its physical constraints. That is why Solo is not a single monolithic model, but a system composed of subsystem models and a higher-level orchestration layer. 

Each model can reason over a bounded domain and produce information about what is happening within that part of the spacecraft. The orchestration layer can then reason across those model outputs to understand the state of the spacecraft and determine what should happen next. The problem is still complex, but the dimensionality of the problem is reduced, and limits Solo to a bounded set of actions. And because we build and test the hardware ourselves, its models can be trained on specialized data generated by the physical systems they will actually operate.

Critically, Solo’s reasoning does not translate into unconstrained control of the spacecraft. Solo’s actions are subjected to a defined command set and rules-based logic that constrains the actions available to the system and governs how its decisions are translated into commands. Solo is this combination of intelligence where the state space is too complex to enumerate with rules alone, and deterministic control where absolute predictability matters most.

Think of it like a map. A map of the entire world is incredibly powerful and useful, but capturing every inch of the planet in the highest possible resolution would mean generating a map the size of the world itself. Most of that information would be useless if you only needed to navigate a small piece of terrain. 

AI has a similar cost. More breadth in a model requires more data and more compute. Solo only needs to navigate one bounded physical system, so there is little value in paying the computational costs of training it on everything else. 

Instead, we want an extraordinarily high-resolution map of the terrain we are actually navigating. For Solo, that terrain is the spacecraft itself.

What specialization enables

Bounding the problem fundamentally changes how we can build the system and what it can do. 

First, we can train models on the hardware they will operate. Before a spacecraft is ever installed on the launch vehicle, all of its components are rigorously tested against the conditions we expect them to encounter during their mission. We can inject anomalies or off-nominal conditions to better understand how the hardware actually behaves across different conditions. That testing generates exactly the kind of data Solo needs. Because we build the spacecraft, we also build the training environment, which turns the spacecraft development process itself into a data engine for the intelligence platform.

Second, we can test more of the relevant state space. As the dimensionality of a model’s inputs and the number of possible states increase, the amount of training data required to cover the permutations also increases. Eventually, a model will encounter combinations of conditions that weren’t represented in its training data. The larger and less constrained that state space becomes, the harder it is to characterize in advance how the system will behave across it. 

Solo deliberately reduces that dimensionality. The more tightly we can define what the model needs to understand, the more precisely we can define what it should reason over. The models don’t need to reason over an open-ended physical world; they can focus on telemetry, states, relationships, and other data that is actually relevant to the spacecraft, and choose from a defined set of actions. 

This constraint in turn makes the model more verifiable, because we have a clearly defined universe of behaviors and outcomes that we can test before handing it control of a spacecraft. 

And third, specialization makes the architecture modular and adaptable. That means adapting Solo to customer hardware doesn’t require us to rebuild the intelligence system from scratch. Instead, relevant models can be specialized around the new hardware via the data generated through testing. 

The future of deep space

This doesn’t mean autonomy in deep space will be easy. A spacecraft is a complex, interconnected physical system that must operate in one of the harshest environments imaginable. Subsystems can interact in ways that are difficult to anticipate in advance, and the number of possible states is enormous. But that complexity is knowable; and by bounding the problem, we make that complexity tractable.

We believe the next major wave of AI will happen in the physical world. Models won’t just generate information; they will reason about physical systems, make decisions, and act on those decisions through hardware. The first phase in this revolution won’t begin by solving the entire physical world at once. The first phase will begin with bounded physical systems that are already rigorously tested, governed by physics, and operated through a defined set of rules and commands. 

For us, that system is a spacecraft.

That’s what we’re building with Solo.