Move any object in space, without launching more.
Stelaris builds swarms of small autonomous robots that grab, move, and place any object in orbit, on demand. It is the logistics layer the space economy is missing. And we build it AI-native: the company runs on AIOS, an operating system where LLM agents do the research and engineering inside one shared context, so a small team ships at a pace the old space industry can’t.
Launch got cheap. Doing anything up there didn’t.
The cost of reaching orbit has fallen roughly 36×, and keeps falling toward ~$100/kg. The cost of moving, servicing, or recovering what’s already up there hasn’t budged, because every job today needs its own bespoke spacecraft. That gap is the opening: a shared, reusable swarm that makes every in-space action cheap.
No single robot could move it. Together, they can.
A swarm converges on an object it has never seen, grabs it, and repositions it with coordinated ion thrust, then lets go and moves to the next job. No propulsion on the payload. The intelligence lives in the swarm, not the cargo.
Current research
● in flightThe head predicts the world well, but a head that predicts isn’t yet a head the policy uses. Whether prediction improves task success is the live question, and three levers are running now. Each is judged only by a prediction-ablation on matched task success, never saliency alone.
Let the value loss shape the head (drop the frozen-feature detach). Does a head the policy can influence become a head the policy actually uses?
Predict progress-to-goal, the task-relevant quantity, instead of raw object motion. Does a more useful target lift coordination?
A neighbor’s force correlates ~0.33 with reward-to-go, versus ~0.00 for heading. Is force the thing worth anticipating?
What working looks like
real runsEvery curve below is a real training run, pulled straight from our experiment logs, not an illustration.
Coordination is the moat.
Launch is nearly solved; thrust is a commodity. The hard problem is getting many autonomous units to agree, in real time and from only what each can sense locally, on how to grip, balance, and steer a shared mass none could move alone. Whoever solves that owns the layer the entire space economy runs on. It is exactly the problem our founder spent his PhD on.
vs. one big servicer
Incumbents build a single bespoke vehicle per target. A swarm handles large, uncooperative, or many-at-once jobs no monolith can, and reuses the same fleet across missions.
Only local information
Each robot coordinates from what it senses nearby: no central controller, O(1) communication, and resilient to losing any single unit.
Scales & transfers
Policies trained on four robots generalize to eight zero-shot, and from simulation to hardware. The coordination layer compounds with every mission.
The company runs on an operating system for AI agents.
Space is hard and great teams are small. So Stelaris runs on AIOS, an operating system that LLM agents work through. They read the research, write and test the code, run experiments, and re-check their own results, all inside one shared, hierarchical context that points them at the goal. Add more agents or swap in a better model, and the whole company gets faster. That is how a handful of people execute like a team many times their size, and it is domain-agnostic, not tied to space.
Shared hierarchical context
Every agent works from one layered source of truth: direction and goals at the top, live state below. A new agent, or a smarter model dropped into any role, inherits the full picture instead of rebuilding it.
Plug in at any layer
A person or an agent joins the distributed fleet at any level, declares intent, and deconflicts before acting. Dozens work in parallel with no central bottleneck and nothing to collide over.
Scales with LLM usage
Any model fills any role behind one swappable seam. Add agents or adopt a better model, and velocity compounds: the company scales with LLM usage, not headcount.
A two-stage plan to the logistics layer.
Stage one proves autonomous multi-robot coordination and a first servicing or relocation mission: the capability no incumbent has. Stage two scales it into servicing and debris-sweep constellations that earn recurring revenue from relocation and the deorbit mandates now written into law.
Coordination + first mission
Demonstrate autonomous multi-robot coordination and a first servicing / relocation mission.
Servicing at scale
Fleet servicing and debris-sweep constellations against the deorbit mandates.
The logistics layer
In-space assembly of large structures as launch nears ~$100/kg.
Built by someone who’s spent his career on this exact problem.
Joshua Bloom holds a PhD in Robotics Engineering (WPI) with a dissertation on Global State Prediction, the coordination method Stelaris runs on. He leads the Autonomy Behaviors & Agents team at Applied Intuition, after engineering multi-agent RL for 300+ agents at MIT Lincoln Laboratory.
Full background & research →Let’s talk.
Stelaris is raising to build the coordination layer of the space economy. If you back deep tech at the frontier, we’d like to show you the science up close: the live research cluster on the right is running right now.