Distributed Observation Scheduling and Sensor Fusion For Large Agile Constellations

• The distributed approach is motivated by robustness, continuous operation without ground control, and handling asset losses through space-to-space decisions.The paper contrasts this motivation with the single-point failure mode of ground-based control.Source: INTRODUCTION · p. 1
• The approach combines decentralized auction-based target allocation and scheduling with distributed data association and fusion.The tasking component distributes target ownership among spacecraft teams, while the fusion component decentralizes target tracking across the constellation.Source: INTRODUCTION · pp. 1–2
• Each tasking team has one command agent and a unique subset of sensing spacecraft, limiting onboard scheduling to a smaller target set.A lightweight scheduler scores auction bids to approximate feasible observation schedules at relatively low computational expense.Source: INTRODUCTION · p. 2
• Novel local information is spatially shared so fusion satellites can combine measurements of the same target, while existing-track information is routed to its owner.Track information is then disseminated through the constellation for data association.Source: INTRODUCTION · p. 2
• The paper presents a simulation architecture and uses it to evaluate decentralized observation scheduling and sensor fusion in a large-scale constellation.The architecture models the distributed systems under study and supports experimental results for the scheduling and fusion algorithms.Source: INTRODUCTION · p. 2

System Overview

• The reference system is a 108-satellite LEO Walker-Delta constellation with inter-satellite mesh links and shared sensing and fusion responsibilities.A centralized ground asset collects and processes observations for centralized estimation, but passively observes tracks in decentralized scenarios.Source: System Overview · p. 2

Decentralized Task Allocation

• Decentralized satellite task allocation includes market-based and optimization-based approaches, with market methods prominent for their implementation simplicity and auction-scoring options.Heuristic geometry-based scores can be computationally inexpensive but may lack fidelity in observation-schedule feasibility, motivating a lightweight scheduler.Source: Decentralized Task Allocation · pp. 2–3

Observation Scheduling

• The observation scheduler uses a lower-fidelity, more discretized MILP approach that selects among precomputed pointing directions.It retains finite sensor viewing capabilities and relative sensing-opportunity performance, while omitting full dynamics.Source: Observation Scheduling · p. 3

Distributed Data Fusion

• Distributed multi-target tracking must create consistent system tracks despite rumor propagation, duplicate information, and locally formed redundant tracks.Track-to-track fusion identifies and resolves redundant tracks; covariance intersection and replaying measurement histories are described as options.Source: Distributed Data Fusion · pp. 3–4

Constellation-Based Target Tracking

• Prior constellation tracking work includes consensus-based Kalman filtering and federated architectures, while onboard fusion is motivated by resiliency and responsiveness.Onboard methods also face SWAP constraints on computation.Source: Constellation-Based Target Tracking · p. 4

Simulation Architecture

• The discrete-event simulation gives each satellite and ground station its own process, models communication delays, and profiles event handlers to assign simulated processing times.A centralized coordinator processes events chronologically; the timing model uses a linear fit to measured handler durations.Source: Simulation Architecture · p. 5

Decentralized Observation Task Allocation

• Command agents schedule observations for their sensing teams and communicate to deconflict task ownership; teams are formed by dividing each orbital plane into two parts using alternating selection.The team formation is intended to maximize continuous target coverage per team while providing overlap between teams in the same plane.Source: Decentralized Observation Task… · pp. 5–6

Tasking Preprocessing

• Preprocessing discretizes satellite pointing options, checks target visibility across satellite-time-target-option combinations, and filters each commander’s targets to those visible to a team member.Visibility is determined by comparing FOV surface-normal orientations with the satellite-to-target vector.Source: Tasking Preprocessing · p. 7

Lightweight Observation Task Scheduling

• A mixed integer linear program scores task bundles by constructing observation schedules that maximize target observations with diminishing returns.The model takes tasks, satellites, time steps, and pointing options as input and returns a schedule and score; each spacecraft can point in only one direction per time step.Source: Lightweight Observation Task Scheduling · pp. 8–9

Bid Warped Consensus Based Bundle Algorithm

• BWCBBA builds task bundles greedily, scores tasks by marginal schedule-score change, communicates bids, and resolves assignment conflicts through consensus.Because MILP-derived scores are not strictly submodular, BWCBBA warps them to maintain consensus properties; iterations stop after a message-free iteration or an iteration limit.Source: Bid Warped Consensus Based Bundle… · pp. 9–10

Distributed Data Fusion

• The fusion architecture compares centralized ground estimation with decentralized estimation in which sensing satellites share measurements and fusion satellites process and reconcile tracks.In the centralized approach, sensing satellites initialize and update monotracks before ferrying associated measurements to ground for stereotrack estimation; in the decentralized approach, measurements pass through gates or are geocast, and fusion satellites reconcile duplicates and plan track handoffs.Source: Distributed Data Fusion · pp. 10–11

Tasking Performance

• Centralized tasking reaches roughly 50% of maximum observation utilization, versus roughly 25% for decentralized tasking; both decline as targets disperse.Despite the utilization gap, both approaches provide similar coverage, with most targets observed at each timestep; incomplete edge-of-horizon coverage can put measured coverage below the lower bound.Source: Figure 8 · p. 13, Tasking Performance · p. 13, Figure 9 · p. 14
• The 108-sensing-spacecraft scenario has a maximum of 64 instantaneously observable targets, with clustered targets disadvantaging single-target ownership.The clustering produces spatially unequal sensing demand, so neighboring teams can lose observation opportunities.Source: Tasking Performance · p. 13
• Conflicts settle at zero around 57 iterations, while mean bids rise and are strongly negatively correlated with conflicts.Message exchange peaks at roughly 15 seconds, when conflicts are declining most steeply.Source: Tasking Performance · pp. 13–14, Figure 10 · p. 15, Figure 13 · p. 15
• 37.5312 is the centralized score; decentralized constellation performance remains within 1% of it across BWCBBA iterations.The authors characterize the performance loss from decentralization and single-target ownership as negligible in this challenging scenario.Source: Tasking Performance · p. 14, Figure 12 · p. 15

Fusion Performance

• For limited-FOV scenarios, decentralized fusion nearly matches the fusion baseline, while full system decentralization degrades performance.The fusion-only scenario’s expanded field of view makes it an easier geometry comparison.Source: Fusion Performance · p. 14, Table 1 · p. 15
• Stereo coverage generally matches stereo observability; all scenarios maintain consistently low RMSE, with the worst values after the second target wave.Stereo coverage lags observability while new tracks initialize, especially after the second wave.Source: Figure 14 · p. 16, Fusion Performance · p. 16, Figure 15 · p. 17
• Despite lower coverage, decentralized fusion with tasking generates more tracks than fusion-only, while centralized fusion creates the fewest.The authors suggest the decentralized tasking scenarios struggled to maintain stable, long-duration target tracks.Source: Fusion Performance · p. 17
• Centralized fusion produces fewer duplicate tracks and resolves them faster; decentralized fusion produces more duplicates and resolves them more slowly.The association ratio remains high across methods but drops after the second target wave; centralized fusion also has higher correlation correctness.Source: Fusion Performance · pp. 16–17, Figure 16 · p. 18

CONCLUSION

• The end-to-end decentralized system degrades compared with a full-FOV system but performs relatively comparably to cases where observation scheduling limits sensing.The comparison concerns a system that allocates and tracks targets.Source: CONCLUSION · p. 18
• The work presents a simulation architecture for investigating distributed tasking and sensor fusion in a large constellation.Source: CONCLUSION · p. 18
• Further work will address duplicate tracks and the system’s robustness in identifying and resolving them.Source: CONCLUSION · p. 18
• The authors plan to investigate mitigating performance losses from single-target ownership through shared ownership and early exit criteria from consensus.Source: CONCLUSION · p. 18
• Future research will explore low-latency responsiveness and potential resiliency to failures in the decentralized approach.Source: CONCLUSION · p. 18