Session
Advanced Technologies 3
Location
Salt Palace Convention Center, Salt Lake City, UT
Abstract
Dynamic tasking for agile Earth-observing satellites (AEOS) adapts preplanned schedules using onboard lookahead sensors to detect cloud cover that obscures targets. Existing approaches decompose tasking into sequential lookahead, maneuver, and capture stages, limiting their ability to manage tradeoffs over long time horizons. Instead, we formulate dynamic tasking as a constrained partially observable Markov decision process (CPOMDP) and introduce Cloud-Aware Dynamic Earth-observation Tasking (CADET), a deep reinforcement learning (DRL) framework that jointly selects maneuvering and sensing actions under resource constraints. We derive a closed-form model for the probability that a target is visible given a spatially aggregated lookahead observation, and we introduce CADET-PLAN, a hybrid variant that embeds a classical spacecraft scheduling solver within the learned policy. We compare CADET and CADET-PLAN in a simulated environment against two classical task schedulers: one with no cloud knowledge and an oracle with perfect cloud knowledge. Across twelve configurations of power budget and lookahead field of view, CADET and CADET-PLAN respectively close 42% and 56% of the performance gap relative to these baselines, while satisfying energy constraints.
Document Type
Event
Energy-Aware Dynamic Tasking for Earth Observing Satellites With Deep Reinforcement Learning
Salt Palace Convention Center, Salt Lake City, UT
Dynamic tasking for agile Earth-observing satellites (AEOS) adapts preplanned schedules using onboard lookahead sensors to detect cloud cover that obscures targets. Existing approaches decompose tasking into sequential lookahead, maneuver, and capture stages, limiting their ability to manage tradeoffs over long time horizons. Instead, we formulate dynamic tasking as a constrained partially observable Markov decision process (CPOMDP) and introduce Cloud-Aware Dynamic Earth-observation Tasking (CADET), a deep reinforcement learning (DRL) framework that jointly selects maneuvering and sensing actions under resource constraints. We derive a closed-form model for the probability that a target is visible given a spatially aggregated lookahead observation, and we introduce CADET-PLAN, a hybrid variant that embeds a classical spacecraft scheduling solver within the learned policy. We compare CADET and CADET-PLAN in a simulated environment against two classical task schedulers: one with no cloud knowledge and an oracle with perfect cloud knowledge. Across twelve configurations of power budget and lookahead field of view, CADET and CADET-PLAN respectively close 42% and 56% of the performance gap relative to these baselines, while satisfying energy constraints.
