Persistent custody of moving ground targets is comparatively easy when every satellite has clean sensors and global team state. It is not easy under range-limited inter-satellite links (ISL), packet dropout, and slot-level false-positive clutter. In a Keplerian+J2 environment with an honest pointing rule (a tip holds only if the reported contact is on the true target), a memoryless local greedy policy achieves 42.4% mean custody on the HARD preset, and shared-policy PPO saturates near 49%— close to the local true-detection ceiling, not a failed optimizer. A graph-conditioned controller that fuses agreeing neighbor contacts and coasts the last confirmed track recovers 92.2% (heuristic) / 92.4% (hierarchical PPO), matching the privileged geometry reference (91.3%). On the EASY preset (no clutter, global links) local greedy already reaches 93.1%, so the performance gap is a HARDonly phenomenon driven by information and track persistence rather than by the inability of PPO to optimize. The contribution is a cluttered, communication-constrained custody benchmark, anempirical demonstration that local policies collapse to the true-detection ceiling, and a graph-conditioned hierarchical policy (heuristic and learned) that recovers the visibility ceiling.
15 September 2026 in International Journal of Nuclear Security
Optical (EO) collection from low Earth orbit fails when cloud opacity exceeds the sensor threshold, so an EO-only policy wastes access windows under partial cloud. We study a small mixed constellation (3 EO + 2 SAR) tasking three fixed areas of interest (AOIs) under a spatiotemporal cloud field. EO succeeds only when site opacity is below 0.35; SAR succeeds in weather at lower quality weight (0.55). On a MIXED preset, EO-only greedy achieves 3.6 successful collects per episode and 33.3% confirmed AOI coverage, while a phenomenology heuristic that collects with EO when predicted clear and with SAR when cloudy reaches 33.8 collects/episode and 100% coverage—about +30 collects/episode and +67 percentage points in coverage. An ablation shows that removing the discrete tip channel changes throughput by only ∼0.2 collects/episode; the operational content is dual-sensor tasking, not the tip bits. Shared-policy proximal policy optimization (PPO) with local observations and invalid-action masking learns from untrained performance (∼3 collects/episode) to about 42 collects/episode on MIXED, matching or exceeding the heuristic. Under partial cloud, EO-only tasking loses most successful collects; a dual-sensor policy that uses SAR when EO is blocked recovers on the order of +30 collects/episode and full AOI coverage in this simulation. The discrete tip channel is secondary; shared PPO can learn the dual-sensor rule. Results are from a controlled simulation, not flight data.
Maintaining persistent custody of dynamic ground targets using constellations of low Earth orbit (LEO) satellites is a critical capability for intelligence, surveillance, and reconnaissance (ISR) missions. Building upon our prior work using centralized PPO with stable-baselines3, this study presents an enhanced multi-agent formulation using PettingZoo ParallelEnv and Ray RLlib with a shared policy architecture. Key improvements include a larger effective field-of-view, slower target dynamics, richer per-agent observations incorporating tip density and velocity cues, and a refined reward structure that strongly incentivizes proactive tipping-and-cueing. The trained policy achieved 71.3% mean custody coverage over 500-step episodes, substantially outperforming random (28.6%) and greedy (38.1%) baselines. Analysis of handoff frequency and per-target performance demonstrates emergent cooperative behavior. These results highlight the value of modern multi-agent RL tooling for space domain awareness applications and provide a reproducible benchmark environment for future research.