Daniel Reynolds, Massachusetts Institute of Technology; Kerri Cahoy, Massachusetts Institute of Technology ; Olivier de Weck, Massachusetts Institute of Technology
https://doi.org/10.64861/BZJX6025
Keywords: Fallback autonomy, Space domain awareness, Space situational awareness, Mega-constellations, Distributed autonomy, Mean field game theory, Model predictive control, Decentralized task allocation, Resilient space architectures, Contested space operations
Abstract:
Future space domain awareness (SDA) and space situational awareness (SSA) missions increasingly rely on large, heterogeneous satellite constellations operating across multiple orbital regimes. While these architectures can provide persistent global coverage and rapid responsiveness, they remain vulnerable to disruptions that degrade or sever centralized command-and-control (C2) and global state awareness. Traditional mission management approaches assume persistent connectivity to centralized planners, synchronized task allocation, and continuous operator oversight. These are assumptions that do not hold in contested or connectivity-denied environments. This paper presents a resilient fallback autonomy architecture designed to sustain coordinated satellite tasking and mission continuity when centralized control becomes unavailable.
The proposed framework introduces a dual-stack autonomy architecture that integrates Mean Field Game Theory (MFGT) and Model Predictive Control (MPC) as complementary decision layers for large-scale satellite constellations comprising hundreds to thousands of heterogeneous agents operating across multiple orbital regimes. At the coordination layer, MFGT enables scalable, distributed task allocation by broadcasting regionally scoped incentive fields that encode evolving mission priorities, task densities, and agent scarcity. These incentive fields function as a lightweight, one-way coordination mechanism, allowing individual satellites to infer evolving global mission needs without requiring persistent peer-to-peer communication or synchronized global views. At the execution layer, MPC refines each satellite’s decisions by generating predictive, constraint-aware tasking plans that respect spacecraft-specific limitations, including power availability, onboard storage, and communication access. Together, these layers enable distributed optimization that remains computationally tractable as constellation size scales, while maintaining physical and operational feasibility.
A simulation campaign was conducted to evaluate this architecture in a heterogeneous AMOS constellation composed of 1,000 satellites distributed across low Earth orbit (LEO), medium Earth orbit (MEO), and geostationary orbit (GEO) regimes. The operating environment was discretized into regional coverage cells encoding evolving task density, priority, service demand, and degraded-awareness state, while each spacecraft was modeled with time-varying resource states, including power availability, onboard storage, queue capacity, and communication accessibility through a finite-capacity ground network. In the assessed simulation, the constellation sustained concurrent sensing, communications, hybrid, and resource-support mission threads while operating through coupled stress conditions rather than a single isolated disruption. These stressors included a solar/RF capacity-degradation sequence with constellation and ground-segment contraction, a five-wave emergency regional-demand campaign, and rolling stale-awareness windows that degraded shared belief freshness while preserving physical routing. Performance was evaluated using task completion rate, priority satisfaction index, priority-weighted throughput, queue burden, emergency-demand completion, top-priority value capture, and regional service continuity relative to a hard-feasible local greedy baseline operating in the same scenario. Although the greedy baseline completed more raw tasks, the MFGT+MPC architecture delivered higher mission value, increasing completed priority-weighted value from 294.23 to 329.67 and raising the priority satisfaction index from 0.4756 to 0.5328. During the emergency-demand campaign, agents autonomously detected and executed urgent regional demands, completing 200 of 201 injected tasks with zero deadline misses, compared with 67 completions under local feasibility alone. Across the full campaign, the architecture preserved 91.07% of top-decile task value, compared with 38.43% for the baseline, demonstrating priority-aware degradation and mission-directed continuity rather than uniform capability collapse.
The results demonstrate that the proposed architecture preserves mission continuity under extreme degradation by enabling self-organizing behavior at the constellation level. Rather than exhibiting uniform graceful degradation, the system transitions into a priority-concentrated operating regime in which surviving agents absorb workload, rebalance regional coverage, and sustain high-value mission threads. This behavior emerges without explicit failure-mode logic, pre-scripted contingency plans, or continuous operator input. Broadcast incentive fields allow the constellation to remain loosely coordinated even under intermittent or asymmetric connectivity, while MPC ensures that individual actions remain resource-feasible and dynamically consistent.
From an SSA/SDA perspective, this work establishes a quantitatively validated alternative to fully centralized mission management for large-scale space architectures. Compared against centralized allocation benchmarks, the framework preserves high-priority task completion and regional continuity under severe asset loss while reducing architectural fragility associated with global-state dependencies. Centralized C2 in this paradigm shifts from continuous task allocator to low-bandwidth intent broadcaster, propagating regional priority information through mean-field incentive signals rather than synchronized task directives. By structurally separating coordination (mean-field propagation) from execution (local predictive optimization), the architecture removes centralized computational bottlenecks while retaining global mission alignment. The present validation targets constellations on the order of tens to hundreds of heterogeneous satellites, with scalability to larger populations governed primarily by broadcast update rates and per-agent MPC solve complexity rather than centralized planner limitations. These findings suggest that distributed fallback autonomy is not merely an enhancement to centralized control, but a structural requirement for resilient next-generation SSA/SDA constellations operating under degraded or connectivity-limited conditions.
Date of Conference: September 15-18, 2026
Track: Space Domain Awareness