Evidence
Every number on this page is produced by scripts in this repo (python -m src.cli all) and committed under artifacts/.
1. Data pipeline: 7-day TLE window → super dataset
| Raw lines / element sets parsed | 97,299 / 32,433 |
| Exact duplicates · same-epoch re-publications · outside window | 0 · 0 · 234 |
| Element sets in history | 32,199 |
| Unique objects (super dataset) | 32,199 |
| Objects with ≥2 elsets (usable for back-test) | 0 |
| Regimes | LEO 27,849 GEO 1,721 HEO 1,617 MEO 1,012 |
| Object types (SATCAT join) | PAYLOAD 20,075 · DEBRIS 9,882 · ROCKET BODY 2,194 · UNKNOWN 48 · |
| Elset discontinuities / maneuvering payloads | 0 / 0 |
| Stale (>3 d) · decaying (perigee <200 km) | 1023 · 61 |
2. How accurate are the predictions? SGP4 back-test
25,840 prediction pairs: older elset propagated to the newer elset's epoch (503 objects with discontinuities excluded).
| LEO horizon | n | median km | p90 km | p95 km | |radial| | |along| | |cross| |
|---|---|---|---|---|---|---|---|
| 0-6 h | 4,617 | 0.108 | 3.71 | 7.356 | 0.011 | 0.097 | 0.013 |
| 6-12 h | 10,478 | 0.178 | 4.951 | 9.397 | 0.025 | 0.17 | 0.017 |
| 12-24 h | 8,489 | 0.59 | 13.922 | 20.086 | 0.045 | 0.579 | 0.037 |
| 24-48 h | 210 | 0.87 | 11.412 | 20.253 | 0.073 | 0.865 | 0.053 |
Fitted per-axis uncertainty (used as debris covariance for Pc): σ = 0.05 km + 0.6074 km/day × elset age.
3. Learning: Double DQN vs vanilla DQN
4. Held-out evaluation (200 scenarios, 6 unseen missions)
156 of 200 scenarios have at least one RED conjunction if nothing is done.
| Policy | Safe (no RED) | 95% CI | Clean | Δv m/s | Slip min | Cost | Screenings | ms / plan |
|---|---|---|---|---|---|---|---|---|
| No avoidance | 22.0% | 16.8–28.2 | 11.5% | 0.0 | 0.0 | 5.95 | 1.0 | 16.1 |
| Random | 73.0% | 66.5–78.7 | 35.0% | 1.173 | 0.505 | 0.751 | 4.93 | 78.3 |
| Greedy lookahead | 100.0% | 98.1–100.0 | 93.0% | 0.581 | 0.605 | -2.248 | 13.71 | 242.5 |
| DQN (vanilla) | 100.0% | 98.1–100.0 | 79.5% | 0.915 | 0.71 | -2.071 | 3.59 | 83.7 |
| Double DQN | 99.5% | 97.2–99.9 | 82.0% | 0.747 | 0.875 | -2.06 | 4.25 | 85.2 |
| Double DQN + guard | 100.0% | 98.1–100.0 | 82.0% | 0.742 | 0.875 | -2.094 | 5.33 | 81.6 |
| Grid search (subset) | 100.0% | 95.4–100.0 | 100.0% | 0.573 | 0.725 | -2.284 | 293.77 | 4737.7 |
| Greedy lookahead (same subset) | 100.0% | 95.4–100.0 | 93.8% | 0.536 | 0.6 | -2.261 | 13.57 | 211.6 |
| Double DQN + guard (same subset) | 100.0% | 95.4–100.0 | 82.5% | 0.685 | 0.95 | -2.111 | 5.25 | 67.3 |
Cost = the agent's objective (risk + Δv + schedule), lower is better. DDQN forward pass: 12.9 µs.
5. Real catalog, no injected threats
30 random launches against the real catalog: 0 with a RED conjunction, 11 with an AMBER one; median closest approach 6.1 km.
| Mission | Object | Miss km | Pc | Risk |
|---|---|---|---|---|
| Firefly Alpha 504.2 km · 2025-03-11T21:42 | HORUS 2 (PAYLOAD) | 3.194 | 7.4e-09 | AMBER |
| Vega 576.8 km · 2025-03-11T02:40 | STARLINK-3055 (PAYLOAD) | 4.288 | 3.2e-21 | AMBER |
| Zhuque-2 521.6 km · 2025-03-11T03:58 | OBJECT M (UNKNOWN) | 3.515 | 6.8e-11 | AMBER |
| Falcon 1 542.0 km · 2025-03-12T14:09 | STARLINK-4285 (PAYLOAD) | 4.309 | 3.5e-08 | AMBER |
| Vega 533.7 km · 2025-03-12T06:54 | STARLINK-5430 (PAYLOAD) | 3.634 | 4.5e-08 | AMBER |
| Tianlong-2 560.1 km · 2025-03-12T11:59 | STARLINK-30213 (PAYLOAD) | 4.177 | 2.3e-07 | AMBER |
| Nuri 538.4 km · 2025-03-12T05:20 | STARLINK-1560 (PAYLOAD) | 2.391 | 5.8e-06 | AMBER |
| Zhuque-2 554.3 km · 2025-03-12T09:30 | STARLINK-30477 (PAYLOAD) | 1.918 | 9.7e-06 | AMBER |
| Astra Rocket 3 492.7 km · 2025-03-12T04:04 | YAOGAN-39 02A (PAYLOAD) | 4.879 | 3.0e-11 | AMBER |
| Long March 4C 540.9 km · 2025-03-12T10:45 | STARLINK-2639 (PAYLOAD) | 3.216 | 4.3e-07 | AMBER |
| Nuri 523.5 km · 2025-03-12T00:58 | DMSP 5D-2 F14 DEB (DEBRIS) | 4.032 | 6.4e-08 | AMBER |