Deep-space relays stop forwarding in queue order: Moon and Mars share one buffer, and bundles are picked by freshness

A deep-space relay no longer decides what to forward and what to drop by queue order; it picks by how valuable each bundle is to the downstream mission. Heuristic rules such as FIFO, LIFO, EDF and Freshest-first do not explicitly optimize task utility and often serve heterogeneous planetary sources unfairly. Across 64 common held-out traffic traces and 320 evaluation episodes, masked Lagrangian PPO has lower mean aggregate excess AoI than all four baselines, and lower than FIFO and EDF on every trace.

The relay spends its scarce transmission opportunities on the freshest bundles rather than on moving more data. The baselines deliver more total bits and use more relay–Earth contact capacity than this controller. The controller reaches a bundle delivery ratio of 0.860 and a bit delivery ratio of 0.686 with 10.02 drops, against 16.20 for FIFO, 17.88 for LIFO, 16.91 for EDF and 16.78 for Freshest-first.

The controller casts relay control as a constrained event-driven semi-Markov decision process and solves it with masked Lagrangian PPO. Feasibility masking removes actions whose transmission cannot finish inside the remaining contact window, and Lagrange multipliers are updated by projection from completed episodic constraint returns. The policy observes only relay-available information and the known contact plan, encodes the buffer with permutation-invariant mean–max pooling, and applies the same physical masks in training and execution.

The result covers a Moon–Mars–relay–Earth scenario with a known contact plan, fixed routing and a single controlled relay, where relay–Earth transmission is non-preemptive and cannot be resumed. Evaluation runs in DTNSim, where five controllers share identical traffic realizations, contact opportunities, relay capacity, expiry semantics and physical feasibility rules, and differ only in action selection. The authors propose next steps of component-wise ablation, contact-plan uncertainty, partial-transfer service, multi-relay control, and validation with operational DTN implementations.

The authors state that the results are descriptive statistics over the fixed held-out policy evaluation and do not constitute a claim of statistical significance or robustness to independent retraining. The finite-horizon constraint differs from the conditional long-run freshness-service guarantee, which additionally requires deficit rate stability. Terminal reflected deficits are diagnostic state variables rather than constraint returns, so positive values do not imply a fairness-constraint violation. The study assumes a known contact plan, fixed routing, a single controlled relay, and non-preemptive no-resume bundle transmission.