Adaptive Science Operations in Deep Space Missions Using Offline Belief State Planning

TL;DR

The paper tackles autonomous, uncertainty-aware science operations for deep-space missions plagued by long comm delays and volatile environments. It introduces a POMDP framework with a Bayesian-network observation model to efficiently represent biosignatures, solved offline by SARSOP to produce verifiable instrument-use policies. Compared to the baseline CONOPS, the approach yields up to ~40% reductions in life-detection misclassifications and demonstrates robustness under off-nominal sample accumulation rates. The methodology is modular and generalizable to other missions, with public code and potential extensions to additional subsystems and resource constraints.

Abstract

Deep space missions face extreme communication delays and environmental uncertainty that prevent real-time ground operations. To support autonomous science operations in communication-constrained environments, we present a partially observable Markov decision process (POMDP) framework that adaptively sequences spacecraft science instruments. We integrate a Bayesian network into the POMDP observation space to manage the high-dimensional and uncertain measurements typical of astrobiology missions. This network compactly encodes dependencies among measurements and improves the interpretability and computational tractability of science data. Instrument operation policies are computed offline, allowing resource-aware plans to be generated and thoroughly validated prior to launch. We use the Enceladus Orbilander's proposed Life Detection Suite (LDS) as a case study, demonstrating how Bayesian network structure and reward shaping influence system performance. We compare our method against the mission's baseline Concept of Operations (ConOps), evaluating both misclassification rates and performance in off-nominal sample accumulation scenarios. Our approach reduces sample identification errors by nearly 40%

Figures (5)

• Figure 1: Selected Bayesian network structure. Arrows indicate conditional dependencies, with parent nodes influencing the distributions of their child nodes. Though many more characteristics and dependencies may exist, we present a simplified model for clarity. \ref{['tab:biosignatures']} contains descriptions of all nodes.
• Figure 2: Full set of conditional probability distributions used in constructing the Bayesian network. Future studies may develop data-driven distributions to improve model accuracy. Some distributions exhibit sharper differences between biotic and abiotic samples, making them more discriminative biosignatures.
• Figure 3: Pareto frontier indicating optimal tradeoff between sample false positive and negative rates from a hyperparameter sweep of $\lambda \in [0.7, 1.0]$. Results for both ConOps baseline (red) and SARSOP solutions (blue) are presented, with an inset highlighting the Pareto curve for lower error regions. The inset further details the clustering of SARSOP solutions and illustrates the balance achieved along the frontier.
• Figure 4: Example set of alpha vectors (expected utilities) for each action over belief in $s_L$ at a sample volume $s_V$ = 60% for the $\lambda = 0.72$ generated policy. At each belief, the optimal action is the one with the highest expected utility, or action value, as determined by the SARSOP-generated policy.
• Figure 5: Dominating policy action at each belief state and sample volume for $\lambda = 0.72$ on the left and $\lambda = 0.935$ on the right. Legend shows actions: declare abiotic ($a_8$, red), biotic ($a_9$, blue), accumulation ($a_7$, gray), and instrument actions ($a_1, \ldots, a_6$ in other colors).