{"id":"c36a6e7e-9e2c-457a-b164-d32fc0734def","arxiv_id":"2608.04590","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A CTDE-PPO agent that jointly selects UAV headings and store-carry-forward routing actions improves simulated DTN delivery ratios over PRoPHET, MaxProp, and Q-learning baselines.","lead":"JUROR is a reinforcement-learning system that jointly decides where small UAVs fly and which nodes should forward stored messages. In a simulated Helsinki road network with 70 nodes, it delivers more messages than standard DTN routing protocols across four traffic patterns.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"JUROR's headline gains over PRoPHET and MaxProp may reflect added UAV heading authority rather than the learned joint policy: no baseline combines fixed/heuristic UAV flight with routing, and Table VI flags JUROR rows as peak held-out values versus single-trial baselines.","rationale":"The reader's weakest_assumption correctly identifies the fairness of the baseline comparison as the load-bearing risk. Reading the paper in good faith, JUROR is a coherent CTDE-PPO formulation: the factored POMDP, per-step team reward, contact-limited observations, and optional auxiliary modules are carefully specified, and the internal ablations (base−UAV, deploy-near-real, base+LSTM, base+HGA) provide useful evidence that UAV relays matter and that contact-limited execution retains much of the benefit. However, none of these ablations separates 'having controllable UAVs' from 'learning the control'. The strongest numerical evidence in Table VI uses selected JUROR peak held-out rows against single-trial baselines, which the paper itself labels as not directly comparable. Because the abstract claims specifically that jointly learned UAV flight and routing produce the gains, the missing heuristic/fixed-UAV baseline is load-bearing. The concrete experiment combining PRoPHET/MaxProp with simple heading heuristics would settle whether the learned heading policy is necessary. I therefore agree with the reader's weakest_assumption and see no reason to change the conditional verdict; if the proposed test reveals no added value from learned headings, the verdict should move toward rejection or unverdictable, but that is not established now.","tokens_in":24048,"tokens_out":6309,"duration_ms":75893,"concrete_test":"One decisive check: in the same discrete-time SCF simulator with N_uav=5, r_uav=900 m and the Sec. VI-A settings, run PRoPHET and MaxProp forwarding under two non-learned UAV heading policies: (i) fix headings to a rotating compass sweep, and (ii) choose each UAV's heading greedily as argmax_k S_{u,k}^{(t)} using the sector-stress vector of Eq. (14). Report results over 10 seeds with means and 95% confidence intervals. If either heuristic-UAV baseline delivers within 5% of JUROR's held-out delivery ratios on M1–M4, the joint learning contribution is not supported; if JUROR's margin exceeds baseline variance, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"For the central claim in the abstract—effective gains over PRoPHET and MaxProp from joint UAV flight and routing—to hold, the improvement must be attributable to the jointly learned policy and not merely to granting JUROR an extra control channel. The comparison in Sec. VI-C does not establish this: every baseline is 'routing-only', and the paper's own descriptions state that PRoPHET, MaxProp, Fan DPUVR, ICC Q-learning, and ICC FQLRP are partially reimplemented with the transceiver stack omitted and with no controllable UAV headings. Table VI marks the JUROR rows as selected peak held-out references 'not directly comparable to the routing-only protocol', and each baseline entry is a single trial with no error bars. The internal base−UAV ablation (Table IV) shows that most of JUROR's edge vanishes without UAVs (M1 peak 13.2 vs. 57.6), demonstrating the value of aerial relays but not of learned heading control: a simple stress-following or fixed-waypoint UAV policy combined with PRoPHET or MaxProp is never tested. Until such a baseline exists, the claimed gain from joint learning is confounded by control authority.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes JUROR, a CTDE–PPO framework that jointly learns per-node opportunistic forwarding decisions and discrete UAV heading actions in a simulated delay-tolerant network. The system model is a factored partially observable MDP with a hand-crafted per-step team reward, decentralized actors, a centralized critic, and optional LSTM-based hotspot prediction and HGA shaping. Experiments on four traffic modes compare JUROR against PRoPHET, MaxProp, Fan DPUVR, and two ICC Q-learning variants, reporting higher held-out delivery ratios for JUROR. The paper also includes ablations over UAV presence, CTDE observation scope, LSTM auxiliary loss, and HGA.","tokens_in":24458,"tokens_out":4208,"duration_ms":48654,"significance":"If the central claim is established, the paper would be a useful contribution to UAV-assisted DTN routing: it formalizes joint motion–routing control as a factored POMDP, preserves decentralized execution, and evaluates on a non-trivial simulator with four traffic modes. The work is also honest about the optional nature of its hotspot auxiliaries, showing that they are traffic-dependent. However, the headline comparison is currently confounded because the baselines are routing-only and cannot control UAV headings, so the reported margin may come from extra control authority rather than from the learned joint policy. The paper's own ablation data show that removing UAVs drastically reduces delivery, which strengthens this concern rather than resolving it. The central claim is defensible but not yet evidenced.","major_comments":[{"comment":"The comparison lacks a control for UAV heading authority. PRoPHET, MaxProp, Fan DPUVR, and the ICC variants are routing-only and do not make UAV heading decisions, while JUROR simultaneously controls both forwarding and discrete headings. Since the base−UAV ablation in Table IV shows that removing UAVs collapses delivery (M1 peak 13.2 vs 57.6), the margin in Table VI could be due to granting JUROR an extra control channel rather than to the learned joint policy. Add baselines that pair PRoPHET and MaxProp with fixed waypoints, random headings, and a simple stress-following heuristic for the same UAV fleet, and report delivery ratios for those controls.","section":"Sec. VI-C, Table VI"},{"comment":"The abstract's claim of 'effective gains over PRoPHET and MaxProp' rests on Table VI, in which the JUROR rows are flagged as selected peak held-out references 'not directly comparable to the routing-only protocol,' while each baseline is a single trial without error bars. Peak-over-training values are not a stable estimator of policy quality. Report terminal or fixed-checkpoint metrics, averaged over multiple training seeds and evaluation episodes, for both JUROR and the baselines under the same episode distribution, and use those numbers for the headline claim.","section":"Sec. VI-A, Sec. VI-C, Table VI"},{"comment":"The baseline implementations are explicitly partial reimplementations: they omit The ONE transceiver stack and substitute simplified proxy metrics, yet no calibration against the original implementations or published results is provided. Lower baseline performance could therefore reflect implementation quality rather than protocol weakness. Validate the reimplementations on a standard benchmark or against published numerical results, and discuss what the omitted transceiver stack could change in the evaluated scenarios.","section":"Sec. VI-C"},{"comment":"The statement that the base−UAV ablation 'verifies joint routing-UA V optimization as the core source of JUROR's performance gains' overreaches: that ablation removes UAVs entirely, so it demonstrates the value of aerial relays, not the value of learned heading control. Separately test a fixed or heuristic heading policy with the same routing mechanisms to support the claimed decomposition.","section":"Sec. VI-D.1"}],"minor_comments":[{"comment":"The sentence introducing JUROR has an unbalanced parenthesis: 'JUROR (Joint UAV flight and Opportunistic Routing, based on the proximal policy optimization (PPO) framework.' should close the first parenthesis.","section":"Abstract"},{"comment":"In the delivery objective, α_s is described as a per-step idle penalty, but as written it is a constant added at every step; clarify how this term incentivizes routing activity and how its magnitude interacts with α_h I_h.","section":"Eq. (18)"},{"comment":"All training runs use seed 42 only; reporting at least a small number of seeds for the default configuration would help assess PPO variance, especially because Table VI reports peak held-out values.","section":"Sec. VI-A"},{"comment":"Reference [19] is listed as an author manuscript without a venue or DOI; if it is not published, mark it as a preprint and consider whether it should be a primary citation.","section":"References"},{"comment":"The traffic-mode description says M2 injects about 4300 messages per episode while Sec. VI-C states 4321; align these numbers for consistency.","section":"Sec. VI-A and Table III"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid formal framework and useful ablations, but the central comparison in the abstract is currently supported only by a comparison the paper itself disclaims. The missing control baseline (routing protocols paired with fixed or heuristic UAV headings) is essential before the 'effective gains over PRoPHET and MaxProp' claim can be evaluated. If the authors add that control and report multi-seed statistics, the paper could become publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: JUROR is a reasonable engineering paper with a genuinely new joint formulation and an honest ablation study, but the central quantitative claim against PRoPHET/MaxProp is not yet supported because no baseline is given control over UAV motion. Peer review should engage with it; the fix is a fair baseline, not a rewrite.\n\nWhat is actually new: the factored CTDE-PPO setup with per-node SCF routing actors and per-UAV discrete heading actors, jointly optimized in one simulator with buffers, TTL, road-constrained ground mobility, and heterogeneous ranges. I do not know of a previous DTN paper that co-learns headings and per-node replication this way. The optional LSTM hotspot predictor and HGA shaping are treated as ablations and disabled by default, which is the right call. The authors also deserve credit for flagging in Table VI that the JUROR peak rows are “not directly comparable to the routing-only protocol” and for admitting the baselines are partial reimplementations. The internal ablation showing base−UAV collapse is meaningful evidence that aerial relays matter, though not that the learned heading policy is the reason.\n\nThe soft spot is the headline comparison. PRoPHET, MaxProp, Fan, and the ICC variants are routing-only; JUROR also controls UAV headings. Without a baseline that combines PRoPHET or MaxProp with fixed waypoints or a simple stress-following UAV policy, the reported margins (0.741 vs 0.505 on M1, and similar elsewhere) could come mostly from the extra control channel, not from the joint learner. The stress-test note lands. On top of that, all numbers are single-trial with no error bars, no code or data are released, and reward-weight sensitivity is not explored, so “peak held-out” is hard to interpret. These are real but fixable weaknesses.\n\nFor a systems/RL reader, the paper is still useful. The problem formulation is careful, the observation design is detailed, and the ablation structure is more honest than most. The right recommendation is to send it to peer review, but with the expectation that the baseline comparison must be reworked and the causal language in the abstract softened until the comparison is clean.","headline":"A plausible joint UAV-routing RL framework whose headline comparison against routing-only baselines is confounded by control authority; worth reviewing, but the abstract needs to be softened until a UAV-enabled baseline is tested.","tokens_in":24909,"tokens_out":1810,"would_cite":false,"duration_ms":24399,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"JUROR's central claim is that routing and UAV flight in a delay-tolerant network should be learned jointly, because each heading choice reshapes the contact graph that forwarding decisions face at the next step.","keywords":["delay-tolerant networks","store-carry-forward","UAV-assisted routing","joint UAV-routing control","centralized training and decentralized execution","proximal policy optimization","opportunistic routing","multi-agent reinforcement learning"],"falsifier":"Run the same simulator with a fixed heuristic UAV flight policy such as heading toward the current congestion centroid while using PRoPHET or MaxProp for forwarding; if that configuration matches JUROR's held-out delivery ratios, then the jointly learned heading-and-routing policy is not the source of the claimed gains.","tokens_in":1657,"feed_emoji":"🚁","tokens_out":1822,"duration_ms":106678,"temperature":0.7,"pith_summary":"The paper is trying to establish that store-carry-forward delivery in sparse, buffer-constrained networks improves when the same learned policy controls both where UAV relays fly and which node forwards which message. It proposes JUROR, a multi-agent reinforcement-learning architecture that trains routing and UAV-heading actors against a single team reward, then lets every node act on only local, contact-limited observations. In simulated Helsinki-medium networks across four traffic modes, JUROR's peak held-out delivery ratios exceed PRoPHET in every mode (e.g., 0.741 vs 0.505 on one-shot single-source bursts and 0.854 vs 0.600 on multi-source bursts), and removing the controllable UAVs collapses performance. If this is right, DTN control should treat mobility as part of the forwarding decision rather than as a fixed input.","feed_headline":"Joint UAV flight and routing beat PRoPHET and MaxProp","feed_subtitle":"One policy chooses UAV headings and forwarding together, lifting peak delivery ratio to 0.85 under bursty traffic.","key_machinery":"The central object is a factored partially observable Markov decision process with sequential motion–routing coupling: at each step the UAVs move first, then routing actions execute under the contact matrix $C^{(t)}$ their motion created, then TTLs decay and contacts rebuild. The mechanism carrying the argument is a shared CTDE–PPO policy with decentralized actors and a training-time centralized critic; each actor outputs a masked choice over $K+1$ forwarding candidates (or idle), and each UAV actor also selects one of eight heading bins, all trained by the per-step team reward. An optional multi-horizon LSTM hotspot predictor and hotspot-guided alignment shaping are auxiliary modules the paper finds traffic-dependent, and the default stack disables them.","core_discovery":"On the paper's own terms, the discovery is that joint UAV–routing control can be cast as a factored, partially observable cooperative decision process and solved under centralized training with decentralized execution. The environment couples the two subsystems in both directions: UAV headings determine which forwarding contacts exist at the next step, and routing outcomes change the buffer-pressure stress fields that steer UAV headings. JUROR factorizes the policy into per-node routing heads and per-UAV heading heads sharing one PPO update, and it trains them with a per-step team reward that rewards deliveries and penalizes expiry, drops, congestion, heading switches, and idle replication. The optional multi-horizon LSTM hotspot predictor and hotspot-guided alignment shaping are auxiliary modules that the paper finds traffic-dependent; the default configuration disables both. The reported result is that this joint policy, evaluated on held-out simulators, delivers substantially more messages than PRoPHET, MaxProp, Fan DPUVR, and ICC Q-learning variants under all four traffic modes, while a deployment-near-real variant using only contact-limited inputs retains most of the peak benefit.","pith_inferences":["An implication the paper leaves implicit is that its reported margins over routing-only baselines may partly reflect the extra control authority JUROR has, since none of the baselines controls UAV headings; rerunning the same simulator with a fixed or heuristic UAV flight policy under PRoPHET or MaxProp routing would isolate the learning contribution.","The evaluation uses one seed, one map, a disk contact model, and hand-tuned reward weights; whether the gains survive across seeds, maps, non-disk radio effects, and reward-weight perturbations is not established by the paper.","The jointness result suggests a broader design principle: in any relay network where nodes control their own trajectories, route utility and motion control should share one objective and one credit signal, so applying the same factored CTDE recipe to surface or underwater mobile relays is a natural next test."],"forward_implications":["UAV trajectory planning should be treated as an action of the forwarding system: every heading choice enlarges or shrinks the set of contacts available to routing, so fixed flight paths leave delivery gains on the table.","Decentralized execution is compatible with the joint policy: the deployment-near-real variant, which uses contact-limited UAV context instead of network-wide statistics, retains most of the peak delivery benefit, so the approach does not require a central controller at run time.","The optional hotspot modules are not always helpful: adding the LSTM auxiliary or HGA shaping improves some bursty multi-source modes but hurts single-source and sustained-congestion modes, which is why the paper's default configuration disables both.","Removing the controllable UAV fleet from the learned policy cuts peak held-out delivery from roughly 57.6 to 13.2 messages on M1, which the paper reads as evidence that joint motion–routing control is a structural prerequisite, not a marginal add-on.","The advantage over routing-only baselines widens under heavier sustained load, where encounter-based scoring methods saturate their buffers."],"supporting_citations":[{"why":"PRoPHET encounter-probability routing; the primary routing-only baseline and the target to beat under all four traffic modes.","marker":"[2]"},{"why":"MaxProp buffer-pressure routing; the second main routing-only baseline, especially relevant under sustained congestion.","marker":"[18]"},{"why":"Fan DPUVR, a hand-crafted UAV-assisted VANET routing protocol used as a baseline and as the representative of decoupled flight and forwarding design.","marker":"[4]"},{"why":"Du et al.'s UAV-assisted vehicular DTN routing, which positions JUROR against prior work that treats trajectories as given.","marker":"[3]"},{"why":"ICC Q-learning and FQLRP, learning-based forwarding baselines that adapt routing but not UAV motion.","marker":"[10]"},{"why":"Proximal policy optimization, the algorithm underlying JUROR's shared actor-critic updates.","marker":"[16]"},{"why":"The ONE simulator, whose native transceiver stack is omitted from the baseline reimplementations, marking the stated limitation of the comparison.","marker":"[29]"},{"why":"GNN-based multiagent DRL for interplanetary networks, used to position JUROR against generic MARL schedulers that lack full store-carry-forward modeling.","marker":"[24]"}],"fun_headline_variants":["Joint UAV flight and routing policy beats DTN baselines","Factored PPO control for UAV-aided delay-tolerant networking","Centralized training, decentralized UAV routing and flight","One policy steers UAVs and forwards messages in DTNs","Joint heading and forwarding control beats PRoPHET and MaxProp"],"cache_read_input_tokens":27008,"weakest_assumption_plain":"The comparison assumes the partially reimplemented routing-only baselines are fair yardsticks, yet none of them controls UAV headings and all omit the full simulator transceiver stack, so part of the margin could come from extra control authority or from implementation differences rather than from the joint learning algorithm.","fun_headline_variants_meta":{"raw":{"variants":["Joint UAV flight and routing policy beats DTN baselines","Factored PPO control for UAV-aided delay-tolerant networking","Centralized training, decentralized UAV routing and flight","One policy steers UAVs and forwards messages in DTNs","Joint heading and forwarding control beats PRoPHET and MaxProp"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000626,"raw_usage":{"total_tokens":2919,"prompt_tokens":988,"completion_tokens":1931,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":1848}},"tokens_in":604,"tokens_out":1931,"duration_ms":16610,"temperature":1.0,"reasoning_tokens":1848,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:18:52.600448+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same simulator with a fixed heuristic UAV flight policy such as heading toward the current congestion centroid while using PRoPHET or MaxProp for forwarding; if that configuration matches JUROR's held-out delivery ratios, then the jointly learned heading-and-routing policy is not the source of the claimed gains.","supporting_citations":[{"cited_title":"Probabilistic Rout- ing in Intermittently Connected Networks,","cited_arxiv_id":null,"evidence_quote":"PRoPHET encounter-probability routing; the primary routing-only baseline and the target to beat under all four traffic modes."},{"cited_title":"MaxProp: Routing for Vehicle-Based Disruption-Tolerant Networks,","cited_arxiv_id":null,"evidence_quote":"MaxProp buffer-pressure routing; the second main routing-only baseline, especially relevant under sustained congestion."},{"cited_title":"A Novel UA V-assisted V ANET Routing Protocol for Post-Disaster Emergency Communica- tions,","cited_arxiv_id":null,"evidence_quote":"Fan DPUVR, a hand-crafted UAV-assisted VANET routing protocol used as a baseline and as the representative of decoupled flight and forwarding design."},{"cited_title":"The ONE Simulator for DTN Protocol Evaluation,","cited_arxiv_id":null,"evidence_quote":"The ONE simulator, whose native transceiver stack is omitted from the baseline reimplementations, marking the stated limitation of the comparison."}],"review_version":1}