02 · System

Three layers bring the service online.

A swarm that moves anything in orbit needs a coordination brain, a body to act with, and a console to command it. We’re building the hardest and most defensible layer first: without the coordination, the robots are just more debris.

01

Algorithms

● Building now

The coordination brain: how a swarm agrees, grips, and maneuvers a shared mass none of its robots could move alone, using only what each unit senses locally. This is the core technical risk, so it’s the piece we’ve built first, in high-fidelity simulation.

Our bet is that prediction is the mechanism behind good coordination. A Global State Prediction head lets each robot anticipate where the payload (and its neighbors) are about to be, and feeds that to the controller choosing actions. It’s the subject of the founder’s doctoral work, and it now runs: the head reaches a within-run prediction correlation around 0.85, reproduced across six independent seeds.

Every claim earns its place by an ablation: disable the component and watch the behavior degrade, or it doesn’t count. Getting there meant first killing the Q-value divergence that used to collapse runs late in training; a stabilized baseline (Huber loss, gradient and target clipping, reward scaling) made results repeatable, without which any comparison is just noise.

490,000+ transport episodes
~1,500 training runs
pred. correlation ~0.85
ablation-gated results

The task we train in

real episode

Four robots, a payload none of them could move alone, a gate they must thread, and a goal on the far side, with nothing but what each robot senses locally. This is collective transport through a gate: the setting where coordination stops being optional.

ConvergeREPLAY · EP - · GSP-N · SEED -
step -real trajectory · 4 robots · 100 kg payload
4 robots, ring-gripped
gate gap 4 m · narrows over a curriculum
goal a 2 m disc · local sensing only
real success · GSP-N seed 42

The gate is the unforgiving test: a constriction the swarm can only clear by collectively aligning and orienting the payload. No single robot can solve it from what it sees. This replay is one real successful episode, positions played straight from the training log.

From this task to orbit

This cylinder-through-a-gate is the lab distillation of the orbital job: the coordination kernel is what ships, and only the setting changes.

The research
Global State Prediction

The founder’s PhD: each agent predicts the collective outcome of the team’s joint action, making a shifting world locally stationary. It beat independent (and even full-information) agents on less bandwidth, transferred sim-to-real on Khepera robots, and scaled 4→8 with no retraining.

The science now
The prediction head, on the gate

We’re stress-testing that idea on the hardest regime (the gate, where coordination is unavoidable) to find which predicted quantity turns the substrate into a measurable gain. That is the live research below.

The company
The same problem, in orbit

Many decentralized units, local sensing, a shared mass none can move alone, threaded through a constraint. The arena becomes 3D orbital dynamics, the cylinder a real vehicle, the robots gecko-pad ion servicers. The coordination problem is identical.

Current research

● in flight

The 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, not by saliency alone.

End-to-end coupling

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?

Goal-progress target

Predict progress-to-goal (the task-relevant quantity) instead of raw object motion. Does a more useful target lift coordination?

Neighbor-force target

A neighbor’s force correlates ~0.33 with reward-to-go, versus ~0.00 for heading. Is force the thing worth anticipating?

Latest runs off the cluster
loading…

What working looks like

real runs

Every curve below is a real training run, pulled straight from our experiment logs, not an illustration.

prediction · mean of 3 seedsinert head
The prediction head came alive. Correlation between the head’s prediction and the true world state, averaged over three seeds (shaded = spread), climbs from noise to ~0.85. An inert run never learns. The signal is real; whether it lifts task success is the open question above.
stabilizeddiverged
Divergence, cured. The same task and seed, two runs. One diverges: value estimates blow up and success collapses to zero (coral). The stabilized learner holds ~0.9 (teal): Huber loss, gradient-clip 10, target-clip 1000, reward-scale 0.1. Repeatability is the precondition for a fair comparison.
success · mean of 3 seeds
Reproducible, not lucky. A baseline learns collective transport and holds it: success climbs and stays high across three independent seeds (shaded = spread). A result worth trusting survives a change of seed.
achievable correlation, by source
The signal was always there. A diagnostic ladder that located the bug behind the head: the online head managed 0.01, but an offline probe of the same inputs reached 0.13, the observable quantity 0.89, and an oracle 0.998: the ceiling was high; the gap was our plumbing, not the task.
02

Hardware

Concept · SV-1

The servicer itself: a flat bus with a gecko-adhesive capture face, twin canted ion thrusters aft, and deployable solar wings, built to dock and thrust in swarms. A first-cut concept grounded in real orbital-servicing hardware.

The choices follow the mission. Gecko adhesion grips objects never designed to be captured, with no docking fixture or cooperative target required. Ion propulsion trades thrust for efficiency, so many small units can make patient, fuel-cheap orbit changes together. Deliberately, the coordination drives the design: the SV-class craft gets designed in simulation first (real geometry, actuation, and sensing) by a dedicated spacecraft designer.

gecko-adhesive capture face
twin canted ion thrusters
deployable solar wings
designed in-sim first
03

Control

Concept preview

The operations layer: one console tasking thousands of robots against thousands of objects: who moves what, when, and along which path. An operator selects a target from live orbital tracking, picks a maneuver (reorient, raise, or de-orbit) and assigns the swarm.

It’s already grounded in real data: the console renders live tracked objects from the public CelesTrak catalog, the same feed behind the tracking globe. That console is the Stage-1 deliverable, an operator sitting down with several assets and several payloads in orbit and tasking the swarm to act on them: a full simulated dry-run of the real product, before any metal is cut.

live CelesTrak catalog
select · plan · assign swarm
reorient / raise / de-orbit
the Stage-1 sim dry-run
See the live training runs, the SV-1 schematic, and the working mission console: Full technical deep-dive →