AstraHeal: Robust Multi-Cycle Autonomous Fault Recovery Under Perturbed Spacecraft Conditions

Autonomous spacecraft operating in Low Earth Orbit (LEO) and deep space must execute closed-loop fault recovery without ground intervention during communication blackouts. While recent aerospace AI architectures demonstrate promising isolated planning and evidential diagnosis, prior studies evaluate autonomous recovery almost exclusively under idealized single-event benchmarks with unperturbed physics models and noiseless telemetry. In operational spaceflight, vehicles encounter sequential cascading anomalies, hardware aging parameter drift, and sensor corruption. In this paper, we evaluate the system-level robustness of the fully integrated AstraHeal architecture—combining Dirichlet evidential uncertainty, counterfactual lookahead planning, and a deterministic Safety Governor—across repeated multi-cycle operations under perturbed conditions. Across eight controlled experimental regimes evaluating 1,320 multi-cycle scenarios and over 18,000 autonomous recovery cycles on a 12U CubeSat Electrical Power System digital twin, we show: (i) AstraHeal isolates and recovers from sequential multi-fault cascades across 3-orbit horizons while delivering 574.0 Wh nominal payload; (ii) system stability is maintained across up to 10 repeated recovery cycles (100% survival across all cycle counts k in {1...10}); (iii) the architecture exhibits bounded graceful degradation under physical parameter perturbations spanning +/-20% in thermal mass, radiator coupling, cell resistance, and solar conversion efficiency (100% survival across 225 perturbed runs); (iv) Dirichlet epistemic uncertainty scales monotonically with sensor noise (sigma = 0.005 to 0.080, climbing from 0.8327 to 0.9988), safely suppressing premature actuations; and (v) exactly zero unsafe action executions (0.00%, exact Clopper-Pearson 95% bound < 0.2791%) occurred across all 1,320 evaluated missions, with 40,711 unsafe proposals deterministically intercepted. In ablation benchmarking, removing counterfactual lookahead planning resulted in a 63.05% drop in delivered mission payload utility (t(49) = 21.09, p = 3.10e-26, Cohen's d = 2.98). These results establish the empirical viability of uncertainty-aware, safety-gated autonomy for long-duration space missions.