A small onboard "AI" scores every tracked object near the satellite for collision risk, the same pattern real missions use: NASA's Perseverance rover runs a terrain-hazard classifier (AutoNav) to pick safe driving paths on its own, JWST's ops software flags anomalous sensor readings automatically, exoplanet-hunting models like ExoMiner classify faint transit dips in starlight, and orbital-debris systems score conjunction risk to schedule avoidance burns.
Risk classifier (logistic model):
risk = ฯ( wโ ยท 1/(d+0.4) + wโ ยท closing_rate โ bias )
ฯ(x) = 1 / (1 + e^-x) (squashes the score into 0..1)
d = distance to satellite, closing_rate = how fast the gap shrinks
bias = 3 โ sensitivity ร 3 (higher sensitivity โ lower bias โ more flags)
Autonomous burn fires when any tracked object's risk > 0.72.
- Tracked-object density โ how many debris fragments are active around the satellite; the classifier only scores active, in-range objects.
- Risk sensitivity โ shifts the classifier's decision bias; higher sensitivity flags marginal conjunctions as risky sooner.
- Sensor range โ objects beyond this radius are invisible to the AI (shown dim grey) and excluded from tracking stats.
- Autonomous avoidance โ when enabled, the satellite fires its own evasive burn the instant a tracked risk score crosses the threshold, without waiting for ground control.