Reconfigurable EOS Scheduling Problem (REOSSP)

• The paper introduces REOSSP, a novel scheduling model coupling constellation reconfiguration through discrete orbital maneuvers with task scheduling to optimize delivered-data performance.
• It formulates a mixed-integer linear program that integrates task sequencing, state propagation, and slot-transition decisions while respecting visibility and maneuver cost constraints.
• Empirical benchmarks indicate that REOSSP outperforms traditional EOS scheduling in data throughput, despite increased computational complexity remedied by rolling-horizon approaches.

Searching arXiv for the primary REOSSP paper and closely related Earth observation scheduling work.

arXiv search query: "ti:\"The Reconfigurable Earth Observation Satellite Scheduling Problem\" OR all:\"REOSSP Earth Observation Satellite Scheduling Problem\""

The Reconfigurable Earth Observation Satellite Scheduling Problem (REOSSP) is an extension of the classical Earth Observation Satellite Scheduling Problem (EOSSP) in which satellites are allowed to perform orbital maneuvers at multiple discrete opportunities to occupy different orbital slots over time, with those slot choices changing future visibility windows and therefore changing feasible observation, downlink, and charging schedules (Pearl et al., 14 Jul 2025). In the formulation introduced under this name, standard EOSSP assumes a fixed constellation and fixed visibility time windows, whereas REOSSP jointly chooses constellation configuration and task execution, using stage-wise orbital maneuvers to alter future target, ground-station, and Sun visibility subject to maneuver-cost and battery limits (Pearl et al., 14 Jul 2025). A plausible implication is that REOSSP should be understood not merely as a richer observation scheduler, but as a coupled constellation-configuration and task-scheduling problem whose feasible task set is endogenous to prior reconfiguration decisions.

1. Definition and conceptual scope

REOSSP is introduced as a new extension of EOSSP motivated by the claim that classical Earth observation scheduling usually assumes a fixed, nadir-pointing constellation with predetermined visibility time windows, while modern satellite systems can instead reconfigure the constellation itself through orbital maneuvers at prescribed opportunities (Pearl et al., 14 Jul 2025). In this setting, each satellite may perform orbital maneuvers at discrete reconfiguration stages to transfer between candidate orbital slots, and the selected slot transitions determine the corresponding visibility matrices for targets, ground stations, and solar charging during each stage (Pearl et al., 14 Jul 2025). REOSSP therefore treats target visibility, ground-station access, and sun exposure not as fixed inputs, but as quantities that depend on orbital-slot choices made during the schedule (Pearl et al., 14 Jul 2025).

This distinguishes REOSSP from several adjacent Earth observation scheduling models. In the exact constraint-programming formulation of the Super-Agile Earth Observation Satellite Imaging Scheduling Problem (SAEOS-ISP), the scheduler chooses among alternative sensing configurations for each strip while accounting for flexible observation windows, variable observation durations, multiple imaging directions, and sequence-dependent transition times, but the problem remains an imaging-layer scheduling model rather than a constellation-reconfiguration model (Caleiras et al., 17 Jan 2026). In agile EOS scheduling with complex-network heuristics, the key reconfiguration-like capability is dynamic attitude change between observations rather than orbital relocation (Wang et al., 2018). In quality-aware deep reinforcement learning for AEOSSP, the decision is a target-time pair under time-dependent profit, but again the platform reconfiguration is attitude and timing rather than orbital-slot transfer (Mercado-Martínez et al., 3 Mar 2025). This suggests that REOSSP, in the strict sense of the term, refers specifically to constellation reconfiguration through orbital maneuvers, not merely agile pointing or mode selection.

The paper defining REOSSP frames the baseline EOSSP as a finite-horizon task scheduling problem in which each satellite may, at each time step, perform one of several mutually exclusive tasks: observe a target, downlink data to a ground station, or charge via solar exposure (Pearl et al., 14 Jul 2025). REOSSP preserves that task structure but enlarges the feasible region by allowing the constellation geometry itself to change over time (Pearl et al., 14 Jul 2025).

2. Mathematical structure

The REOSSP formulation is built on sets of ground stations $\mathcal{G}$, orbital slot options $\mathcal{J}$, satellites $\mathcal{K}$, targets $\mathcal{P}$, reconfiguration stages $\mathcal{S}$, and discrete time steps $\mathcal{T}$ (Pearl et al., 14 Jul 2025). For the full-horizon formulation, the planning horizon is partitioned into equally spaced reconfiguration stages, with $\mathcal{T}^s$ denoting the set of time steps within stage $s$ (Pearl et al., 14 Jul 2025). Each satellite $k$ has a stage-dependent set of candidate orbital slots $\mathcal{J}^{sk}$, and the initial set $\mathcal{J}$0 is a singleton representing its starting slot (Pearl et al., 14 Jul 2025).

The principal REOSSP decision variables are orbital-slot transition variables and task variables. The binary variable

$\mathcal{J}$1

equals 1 if satellite $\mathcal{J}$2 transfers from slot $\mathcal{J}$3 in stage $\mathcal{J}$4 to slot $\mathcal{J}$5 in stage $\mathcal{J}$6 (Pearl et al., 14 Jul 2025). The binary variables

$\mathcal{J}$7

indicate whether, at time $\mathcal{J}$8 in stage $\mathcal{J}$9, satellite $\mathcal{K}$0 observes target $\mathcal{K}$1, downlinks to ground station $\mathcal{K}$2, or charges, respectively (Pearl et al., 14 Jul 2025). Continuous state variables

$\mathcal{K}$3

track onboard data storage and battery level (Pearl et al., 14 Jul 2025).

The REOSSP objective is

$\mathcal{K}$4

with figure of merit

$\mathcal{K}$5

where downlink is weighted more heavily than observation through a positive constant $\mathcal{K}$6 (Pearl et al., 14 Jul 2025). The paper interprets the figure of merit as total amount of data downlinked (Pearl et al., 14 Jul 2025). This makes REOSSP a throughput-oriented scheduling model in which collected data is most valuable once delivered.

A central structural feature is that visibility becomes slot-dependent. Target, ground-station, and Sun visibility are encoded by

$\mathcal{K}$7

which indicate whether target $\mathcal{K}$8, ground station $\mathcal{K}$9, or the Sun is visible at time $\mathcal{P}$0 in stage $\mathcal{P}$1 when satellite $\mathcal{P}$2 occupies orbital slot $\mathcal{P}$3 (Pearl et al., 14 Jul 2025). These parameters enter task-feasibility constraints through the chosen transitions $\mathcal{P}$4, so the scheduler can activate an observation, downlink, or charging action only if the selected slot makes the corresponding object visible (Pearl et al., 14 Jul 2025). This is the defining algebraic coupling between reconfiguration and scheduling.

The maneuver budget is represented by slot-transition costs $\mathcal{P}$5 and per-satellite limits $\mathcal{P}$6, with total transfer cost constrained by

$\mathcal{P}$7

(Pearl et al., 14 Jul 2025). The paper states that maneuver costs are computed analytically using astrodynamics models from Vallado, using circular coplanar phasing formulas for the random instances and seven maneuver types for the Hurricane Sandy case study (Pearl et al., 14 Jul 2025).

3. Operational constraints

REOSSP inherits and extends the operational constraints of EOSSP. At any time step, a satellite may perform at most one power-related task—observation, downlink, or charging—through the exclusivity constraint

$\mathcal{P}$8

for each stage, time step, and satellite (Pearl et al., 14 Jul 2025). Observation is allowed only when the chosen slot makes the target visible; downlink requires ground-station visibility; charging requires Sun visibility (Pearl et al., 14 Jul 2025).

The data-storage dynamics are stage-coupled and time-coupled. Within a stage,

$\mathcal{P}$9

and stage-end states initialize the next stage (Pearl et al., 14 Jul 2025). Storage must remain between $\mathcal{S}$0 and $\mathcal{S}$1 (Pearl et al., 14 Jul 2025). The battery dynamics are similar, but include observation, downlink, time-drift, charging, and reconfiguration consumption. In particular, the stage-boundary battery update includes the maneuver term $\mathcal{S}$2 multiplied by selected slot transfers (Pearl et al., 14 Jul 2025). Battery must remain between $\mathcal{S}$3 and $\mathcal{S}$4, and initial battery is taken as full in the baseline (Pearl et al., 14 Jul 2025).

The model therefore couples three temporal layers: task sequencing inside stages, state propagation across time steps, and orbital-slot path continuity across stages. A plausible implication is that REOSSP combines features of resource-constrained project scheduling, multistage network design, and satellite mission planning, though the paper itself presents it as a mixed-integer linear program (Pearl et al., 14 Jul 2025).

4. Reconfiguration mechanism

In REOSSP, reconfiguration is modeled at discrete, equally spaced stage boundaries, not continuously and not between arbitrary adjacent task executions (Pearl et al., 14 Jul 2025). Within each stage, the orbit or slot is fixed; only at stage boundaries may a satellite move to another candidate orbital slot (Pearl et al., 14 Jul 2025). The chosen slot then determines the target, ground-station, and Sun visibility matrices throughout the following stage (Pearl et al., 14 Jul 2025).

This abstraction is derived from the Multistage Constellation Reconfiguration Problem (MCRP) (Pearl et al., 14 Jul 2025). In the random instances, orbital slots vary only in argument of latitude, i.e., phase (Pearl et al., 14 Jul 2025). In the Hurricane Sandy case study, slots also vary in inclination and right ascension of the ascending node, so the slot library includes both phasing and plane-change options (Pearl et al., 14 Jul 2025). The paper does not embed continuous trajectory optimization inside the MILP; instead, it uses a precomputed slot-transition network with known costs and per-maneuver battery penalties (Pearl et al., 14 Jul 2025). The authors identify this as essentially a high-thrust scope and suggest low-thrust trajectory optimization as future work (Pearl et al., 14 Jul 2025).

The paper also proves that if $\mathcal{S}$5 for all satellites, REOSSP collapses to EOSSP, because only zero-cost stay-in-place transitions remain feasible and the visibility sequence becomes identical to the baseline (Pearl et al., 14 Jul 2025). This establishes REOSSP as a strict generalization of EOSSP.

A useful contrast appears in other Earth observation scheduling formulations. In the super-agile CP model, reconfiguration is represented by flexible observation windows, variable duration, discrete sensing direction, and sequence-dependent transition times

$\mathcal{S}$6

inside a fixed orbital setting (Caleiras et al., 17 Jan 2026). In agile multi-satellite oversubscribed target scheduling, maneuver feasibility is encoded through pairwise mission transition times within an orbit, with energy and memory constraints but no orbital relocation (Wang et al., 2018). In freshness-aware AEOSSP with onboard processing, time-dependent observation value and attitude transition time are the main reconfiguration-like components (Mercado-Martínez et al., 13 Jun 2025). These comparisons indicate that REOSSP broadens the scope of Earth observation scheduling from reconfigurable observation state to reconfigurable constellation geometry.

5. Exact solution and rolling horizon procedure

The defining paper gives a mixed-integer linear programming formulation for both EOSSP and REOSSP and solves them using MATLAB/YALMIP and Gurobi 12.0.0 (Pearl et al., 14 Jul 2025). The authors emphasize that exact REOSSP optimization becomes computationally expensive as the numbers of satellites, stages, and slot options grow, due to the simultaneous presence of binary slot-transition decisions, binary task decisions at every time step, intertemporal data and battery dynamics, and visibility patterns that depend on chosen slots (Pearl et al., 14 Jul 2025).

To address this, the paper develops a rolling horizon procedure (RHP) (Pearl et al., 14 Jul 2025). The full $\mathcal{S}$7-stage REOSSP is decomposed into smaller subproblems RHP$\mathcal{S}$8, where $\mathcal{S}$9 is the current control stage and $\mathcal{T}$0 is the number of lookahead stages (Pearl et al., 14 Jul 2025). Each subproblem optimizes over stages $\mathcal{T}$1, but only the first stage’s decisions are committed (Pearl et al., 14 Jul 2025). Then the horizon rolls forward, and remaining maneuver budget, occupied slot, data state, and battery state are updated using the previously committed solution (Pearl et al., 14 Jul 2025).

The updated maneuver budget is

$\mathcal{T}$2

and analogous update equations are given for inherited slot, data, and battery states (Pearl et al., 14 Jul 2025). In experiments, the paper uses $\mathcal{T}$3, so each subproblem sees the current stage plus one lookahead stage (Pearl et al., 14 Jul 2025).

The RHP preserves much of the benefit of reconfiguration while sharply reducing runtime. However, it generally consumes more maneuver budget than the full REOSSP because the limited lookahead encourages locally effective but globally expensive early moves (Pearl et al., 14 Jul 2025). This interpretation is explicit in both the random benchmarks and the Hurricane Sandy study (Pearl et al., 14 Jul 2025).

6. Empirical behavior and benchmarking

The paper benchmarks EOSSP, REOSSP, and RHP on 24 random instances and one Hurricane Sandy case study (Pearl et al., 14 Jul 2025). In the random instances, the number of stages varies over $\mathcal{T}$4, the number of satellites over $\mathcal{T}$5, and the number of slot options per stage and satellite over $\mathcal{T}$6, with two ground stations, ten targets, a ten-day horizon, and 100-second time steps (Pearl et al., 14 Jul 2025).

Across the random instances, the reported average improvement in objective value is 129.19% for REOSSP over EOSSP and 99.53% for RHP over EOSSP, while RHP is on average 14.74% worse than REOSSP (Pearl et al., 14 Jul 2025). Average runtime is 2.36 min for EOSSP, 29.34 min for REOSSP, and 8.98 min for RHP (Pearl et al., 14 Jul 2025). The full REOSSP failed to find a feasible solution within the 60-minute limit in five larger instances, whereas RHP still produced feasible schedules and outperformed EOSSP in those cases (Pearl et al., 14 Jul 2025). Average total maneuver expenditure is 2.97 km/s for REOSSP and 3.66 km/s for RHP (Pearl et al., 14 Jul 2025).

In the Hurricane Sandy case study, the objective values are $\mathcal{T}$7, $\mathcal{T}$8, and $\mathcal{T}$9, with delivered-data figures of merit $\mathcal{T}^s$0 GB, $\mathcal{T}^s$1 GB, and $\mathcal{T}^s$2 GB (Pearl et al., 14 Jul 2025). Thus REOSSP improves over EOSSP by 288% in objective value and 300% in delivered data, while RHP improves over EOSSP by 192% in objective value and 200% in delivered data (Pearl et al., 14 Jul 2025). Runtime is 0.24 min for EOSSP, 54.60 min for REOSSP, and 6.65 min for RHP (Pearl et al., 14 Jul 2025). Maneuver expenditure is 2.26 km/s for REOSSP and 2.89 km/s for RHP (Pearl et al., 14 Jul 2025).

These results support the paper’s central claim that constellation reconfigurability enlarges future visibility opportunities and can substantially increase delivered-data throughput (Pearl et al., 14 Jul 2025). A plausible implication is that REOSSP is especially valuable when target motion, ground-station access, or regional coverage needs make fixed-geometry visibility the dominant bottleneck.

7. Relations to neighboring research areas

REOSSP sits at the intersection of several strands of research. The exact CP model for SAEOS-ISP shows how flexible observation windows, multiple imaging directions, variable durations, and sequence-dependent slews can be modeled compactly with optional interval variables, sequence variables, alternative, and noOverlap constraints (Caleiras et al., 17 Jan 2026). That work is best viewed as an exact model for a core imaging-layer REOSSP variant, but it does not include constellation reconfiguration (Caleiras et al., 17 Jan 2026).

EOS-Bench provides a benchmark framework for EOS scheduling over 1,390 scenarios and 13,900 instances, including agile and non-agile satellites, energy and storage capacities, transition-time conflicts, and scenario-complexity descriptors such as opportunity density, task flexibility, conflict intensity, and satellite congestion (Yin et al., 28 Apr 2026). The paper explicitly notes that the current release is offline, deterministic, optical-only, and does not include explicit reconfiguration decisions, while identifying “heterogeneous, reconfigurable, and very-large constellations” as future work (Yin et al., 28 Apr 2026). This suggests that EOS-Bench is a practical foundation for benchmarking REOSSP-like variants, even if it is not itself a full REOSSP benchmark.

The market-based and DCOP-based work on EOS constellation scheduling with exclusive orbit portions is relevant to REOSSP because it studies partial ownership, local autonomy, and coordination without full schedule disclosure (Picard, 2021). There, the central planner may place non-exclusive requests inside exclusive windows through auction or DCOP mechanisms (Picard, 2021). This does not constitute constellation reconfiguration, but it addresses a related problem of reallocating effective capacity under decentralized control.

Theoretical scheduling in dynamically reconfigurable optical and datacenter networks provides further context. Online scheduling for weighted flow and completion time in reconfigurable networks studies direct source-destination transfers in a degree-bounded dynamic topology, with competitive guarantees under speed augmentation but no explicit reconfiguration delay (Dinitz et al., 2020). Opportunistic-link scheduling in two-tier reconfigurable datacenters studies online weighted latency minimization through stable matching in a hybrid fixed/reconfigurable network, again under resource augmentation and without explicit setup penalties (Kulkarni et al., 2020). These works are not Earth observation models, but they supply abstraction patterns for reconfigurable scheduling under dynamic topology.

At a semantic level, “glued partial orders” for reconfigurable interaction distinguish ordinary nondeterministic choice from forced interleaving induced by reconfiguration (Alrahman et al., 2021). While this is not a scheduling algorithm, it suggests a conceptual distinction between structural alternatives and schedule refinements that may be relevant when reasoning about REOSSP state-space compression. This suggests that some schedule differences in reconfigurable systems may be artifacts of reconfiguration-induced serialization rather than genuinely distinct operational choices (Alrahman et al., 2021).

8. Limitations and open directions

The REOSSP formulation as introduced is deterministic: target, visibility, and other schedule characteristics are assumed known in advance (Pearl et al., 14 Jul 2025). Orbital maneuvers are represented by precomputed slot transitions rather than embedded continuous trajectory optimization (Pearl et al., 14 Jul 2025). The current scope is effectively high-thrust, and the authors explicitly suggest low-thrust trajectory optimization as future work (Pearl et al., 14 Jul 2025). No stochastic target uncertainty is modeled, despite the relevance of uncertainty in disaster monitoring (Pearl et al., 14 Jul 2025).

The formulation also differs from agile EOS scheduling models that account for off-nadir image degradation, flexible dwell duration, or alternative imaging direction. In REOSSP, the reconfigurable degree of freedom is orbital placement rather than look-angle selection (Pearl et al., 14 Jul 2025). A plausible implication is that a more complete framework would integrate both orbital reconfiguration and agile sensing-layer choice. The defining paper explicitly suggests developing a similarly rich agile EOSSP model for direct comparison with constellation reconfigurability (Pearl et al., 14 Jul 2025).

Benchmarking is another open issue. EOS-Bench is broad and open-source, but it currently lacks explicit reconfiguration states, payload mode switching, orbital transfer decisions, and state carryover across time (Yin et al., 28 Apr 2026). Extending it with configuration-state variables and transition decisions would make it a more natural benchmark platform for REOSSP research (Yin et al., 28 Apr 2026).

A further open direction concerns exact versus approximate optimization. The defining REOSSP paper shows that full-horizon exact optimization yields the strongest performance but often becomes intractable, while RHP improves runtime substantially at the cost of solution quality and greater maneuver expenditure (Pearl et al., 14 Jul 2025). This suggests a natural role for decomposition, approximation, and learning-based methods. Existing Earth observation work already demonstrates strong CP modeling for exact imaging-layer scheduling (Caleiras et al., 17 Jan 2026), graph-based heuristics for agile multi-satellite oversubscription (Wang et al., 2018), quality-aware deep reinforcement learning for target-time selection (Mercado-Martínez et al., 3 Mar 2025), and freshness-aware constructive plus local-search heuristics with onboard processing (Mercado-Martínez et al., 13 Jun 2025). A plausible implication is that future REOSSP solvers may combine exact subproblem models for short-term resource feasibility with heuristic or learned methods for long-horizon reconfiguration planning.

9. Significance

REOSSP formalizes a shift in Earth observation scheduling from choosing tasks within a fixed access pattern to choosing both constellation geometry and tasks within a coupled optimization model (Pearl et al., 14 Jul 2025). In the strict sense established by the naming paper, it is the scheduling problem in which satellites are allowed to perform orbital maneuvers at multiple discrete opportunities to occupy different orbital slots over time, with those slot choices changing future target, ground-station, and Sun visibility (Pearl et al., 14 Jul 2025). The resulting model integrates multistage constellation reconfiguration, observation scheduling, downlink scheduling, charging, data storage, battery dynamics, and maneuver-cost budgeting in a single MILP (Pearl et al., 14 Jul 2025).

Within the broader literature, REOSSP occupies a distinct position. It is narrower than the most general notion of reconfigurable scheduling if one includes arbitrary sensor modes, communication routing, or stochastic replanning, but broader than standard EOSSP and most agile imaging formulations because it makes the orbit configuration itself a decision variable (Pearl et al., 14 Jul 2025). The empirical evidence reported so far indicates that this added decision space can substantially improve delivered-data performance, especially in dynamic monitoring scenarios, while also introducing major computational challenges that motivate rolling-horizon and other approximate solution methods (Pearl et al., 14 Jul 2025).