At 500 kilometres, no one is coming to save you: when autonomy is not optional

There is one detail that changes everything.

If a robot at home stops working, you restart it. You call tech support. In the worst case, you take it in for repair. There's a safety net. But at 500 kilometres above the Earth, orbiting at 28,000 km/h, in an environment where the temperature swings between -170 and +120 degrees depending on whether the sun hits or not, that net disappears. The system decides on its own, or the mission dies.

That is the reality of space robotics. And it is also the reason why autonomy, in this context, is not an optional feature. It is a condition for survival.

The problem no one sees
For decades, the dominant model in space was the monolithic satellite: a large, expensive, highly specialised platform that concentrated all mission capabilities into a single vehicle. That model is changing. The miniaturisation of electronics, reusable launchers and the falling cost of access to space have opened the door to constellations and formations of small satellites that cooperate to do what a single one used to do alone. Initiatives like Starlink, OneWeb or Planet Labs' observation constellations are the visible expression of that shift.

But behind it lies a technical problem that never makes it into press releases: how do you get multiple satellites to maintain a precise, coordinated geometry over time, without a central node controlling everything, under real orbital perturbations that degrade any formation if not actively compensated? Earth's oblateness, atmospheric drag, solar radiation pressure — constant, invisible forces that push and pull each satellite away from its ideal trajectory.

That was exactly the problem I tackled in my thesis.

Learning in simulation, acting in the void
The solution I explored is based on multi-agent reinforcement learning (MARL): several agents that learn to cooperate through experience, without anyone programming explicit behavioural rules. It is the same family of techniques used to teach an AI to play chess or drive a car, but applied to an environment where mistakes cannot be undone and communication with the ground has latency.

The process is not immediate. Training the final policy required approximately 10.75 million simulation steps, split across four progressive phases:

1. A single agent learning to position itself on the target ring.
2. Two agents beginning to coordinate, where the symmetry of the problem allowed a natural distribution to emerge without being explicitly encoded.
3. Three agents facing the geometric ambiguity of deciding who goes where — the hardest phase, where the collision rate reached 8%.
4. Four agents operating simultaneously under real orbital perturbations at 100% nominal magnitude: atmospheric drag, solar radiation pressure and the effect of Earth's oblateness (J2).

The final model, PPO_final, achieved a 99.1% success rate over 1,000 evaluation episodes, with a mean positioning error in the order of 5 millimetres and terminal velocities close to zero. The four agents learned to coordinate, avoid collisions and recover the formation after impulsive perturbations — all without external instructions at runtime.

The problem is that all of it happened inside a computer.

The gap between simulating and existing
There is a concept in robotics that captures one of the field's most honest challenges: the Sim-to-Real gap — the divide between what a system learns in simulation and what it can actually do when it faces the physical world.

In a previous article I reflected on how simulating is not the same as feeling. There is a technical version of that same problem here: no matter how precise the simulator, the real world holds sensor noise you never modelled, tiny frictions that accumulate, delays that the digital environment abstracts away. The question every research effort in this field must answer is how much of what was learned in the computer survives contact with physical reality.

To get closer to that answer, I developed a second line of work: training the Ufactory Lite6 robotic arm on an air-bearing platform that simulates microgravity conditions. The policy is trained in MuJoCo and transferred directly to the real hardware, without modifications. The calibration error between the simulated and real systems was under 2 millimetres in position, with run-to-run reproducibility below 2%. The simulator had captured what mattered.

But it is worth being honest about the limits: the four-satellite formation is still, for now, simulation only. The full leap to real multi-agent hardware is still pending, and that is where things will get truly difficult — and truly interesting.

Why this matters
The reason terrestrial robotics and space robotics are, at their core, the same technological challenge is precisely this: every algorithm we refine down here, in controllable and repairable environments, is a step toward systems that will one day operate where no one can come to rescue them.

The large constellations that today provide internet coverage in remote areas, monitor climate or detect wildfires still depend heavily on instructions from the ground. But as those formations grow in size and complexity, that model has a clear ceiling. Distributed autonomy — each satellite deciding from local information, without needing anyone from below to tell it what to do — is not a future option. It is the inevitable path.

And the open questions are enormous: can real coordination emerge between agents without explicitly encoding how to distribute themselves? What happens when one agent fails and the others must reconfigure the formation without external instructions? How far does a policy trained in simulation hold up when the chaos of the real world truly arrives?

These are questions that the field of autonomous space robotics is still working to answer. And that makes them worth pursuing.