{"id":"c70e90a7-d4f2-47c7-9006-d0df8a166117","arxiv_id":"2509.01427","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An enhanced SAC algorithm with temporal sequence input, layer-normalized GRU, and squeeze-excitation blocks jointly optimizes drone trajectories and communication schedules for distributed-beamforming drone relays, reducing simulated AoI and drone energy.","lead":"Multiple drones cooperatively relay sensor data to a distant base station using distributed beamforming, and a deep reinforcement learning agent, SAC-TLS, decides where they fly and when they transmit. The paper claims this lowers the average age of sensor information by 17.3% and cuts drone energy against several baselines in simulation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"A2A time budget unverified: Eq. (7) computes a nonzero δA2A, yet the lossless-broadcast assumption is never checked against slot length; if δA2A consumes a significant fraction of each slot, the claimed 17.3% AoI gain is not credible.","rationale":"Good-faith reading: the paper's central claim is that SAC-TLS achieves lower time-average AoI and energy than strong baselines in the simulated distributed-beamforming relay system. What must be true is that the simulator's AoI dynamics (Eq. 10) faithfully represent the modeled physical system over the evaluation horizon. The most insecure link is the A2A/broadcast and phase-coherence idealization: the paper itself writes a finite A2A rate (Eq. 6) and a finite broadcast phase (Eq. 7), then assumes lossless short-range broadcast. It also assumes the VAA's SNR is the coherent sum of amplitudes (Eq. 8) without any synchronization cost. If those assumptions fail, Q(t) (Eq. 9) overestimates the fraction of data forwarded, and AoI is understated. The reader's verdict already conditions on this; our stress test adds a concrete quantitative check: compute δA2A from Eq. (7) with the paper's published parameters (or plausible defaults) and verify the per-slot time budget. This is directly testable, and it targets the mechanism behind the 17.3% AoI reduction rather than a generic reproducibility complaint. I do not treat the absence of code as a separate fatal flaw; instead it makes the proposed check necessary. The paper deserves credit for a clearly structured MDP and a plausible DRL architecture; the issue is the environment model, not the learning algorithm itself.","tokens_in":21285,"tokens_out":10877,"duration_ms":131993,"concrete_test":"Re-run the Section 6 simulation with the A2A model activated: instantiate Eq. (6) with plausible parameters (e.g., ρ0 = -60 dB, α=2, P_j'=0.1 W, B_j'=20 MHz, and the D_i values used in Fig. 4), compute δA2A per time slot for the trained SAC-TLS trajectory, and force δA2G = min(δA2G, 1-δG2A-δA2A-δmove). If the average AoI increases by more than 10% relative to the lossless-A2A case, or if SAC-TLS no longer beats SAC/TD3 by the reported margins, the central claim is conditional on an unchecked assumption. As a second check, compare the lossless-A2A assumption to a phase-error model in Eq. (8) with random phase offsets uniformly in [-ε, ε] for ε=5°; if the 17.3% gain does not survive ε>0, the beamforming idealization is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 defines a finite A2A broadcast rate (Eq. 6) and a broadcast phase δA2A(t)=max_j S_j/R_j^{A2A} (Eq. 7), then immediately assumes 'all data collected from SNs can be reliably broadcast ... within each time slot'. That assumption is only valid if δA2A(t) is negligible relative to the unit slot. The paper never reports the A2A parameters needed to check this (B_j', P_j', ρ0, α, D_i) nor verifies δG2A+δA2A+δA2G+δmove ≤ 1 for the trained policy. The same idealization extends to Eq. (8): coherent VAA beamforming requires phase alignment, but no synchronization overhead or phase error is modeled. Both assumptions sit directly upstream of Q(t) in Eq. (9) and the AoI update Eq. (10). Any violation shrinks the A2G capacity and increases AoI for all algorithms, potentially eliminating the reported 17.3% advantage. This is not a minor parameter choice; it is the mechanism by which the system avoids the inter-AAV bottleneck and achieves the claimed freshness gains.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies an IoT data-forwarding system in which a swarm of AAVs forms a virtual antenna array (VAA) to relay sensor data to a remote base station, with the goal of minimizing time-average age of information (AoI) and AAV energy consumption. The problem is formulated as a non-convex mixed-integer program over AAV trajectories and communication schedules, and a deep reinforcement learning algorithm, SAC-TLS, is proposed by augmenting soft actor-critic with temporal sequence input, layer-normalized GRU (LNGRU), and squeeze-and-excitation (SE) blocks. Simulation experiments compare SAC-TLS with greedy, TD3, PPO, TQC, and SAC baselines, and report a 17.3% reduction in average AoI, a 24.5% increase in cumulative reward, 1.6x faster convergence, ablation results, and pruning performance under 90% compression.","tokens_in":21670,"tokens_out":4289,"duration_ms":54348,"significance":"The topic is timely and the proposed algorithmic components (temporal sequence modeling, LNGRU, SE blocks) are plausible enhancements for a dynamic UAV/IoT setting. The paper also provides a complexity analysis and an ablation/pruning study. If the reported gains were supported under realistic communication assumptions, the work would be a useful step toward multi-AAV distributed-beamforming AoI optimization. However, the quantitative claims are currently conditional on a non-standard AoI update, an unverified lossless A2A broadcast assumption, and a simplified MDP that removes the scheduling variable from the learned action space. Thus, the significance is not yet established at the level claimed.","major_comments":[{"comment":"The AoI update A_i(t+1)=(1-Q(t))(A_i(t)+1) when SN i is scheduled is not the conventional AoI definition. Standard AoI after successful delivery resets to the age of the delivered update (typically 1), and partial progress does not continuously scale down the age. Here, Q(t)<1 multiplicatively reduces the age even if only a fraction of the data is forwarded, and Q(t)=1 gives A_i(t+1)=0, which is inconsistent with data generated at the current slot. This non-standard metric is exactly the term minimized in the reward (Eq. (20)), so the reported 17.3% AoI gain in Section 6 may be an artifact of the metric itself. The authors should derive Eq. (10) from a packet/freshness model or cite a precedent; otherwise, they should adopt a standard AoI definition and rerun the evaluation.","section":"Section 3.3, Eq. (10)"},{"comment":"The A2A phase is modeled with a finite broadcast rate and a duration delta_A2A(t)=max_j S_j/R_j^{A2A}, but the text then assumes that all collected data can be reliably broadcast within each time slot. No simulation parameters for the A2A link (B_j', P_j', rho0, alpha) or sensor data sizes D_i are reported, and no verification is given that delta_G2A + delta_A2A + delta_A2G + delta_move <= 1 for the trained policy. In parallel, Eq. (8) assumes perfect coherent beamforming with no synchronization overhead or phase error. Both idealizations sit directly upstream of Q(t) in Eq. (9) and the AoI update Eq. (10). If these phases are not negligible or phase alignment is imperfect, the A2G capacity shrinks and the claimed AoI improvement may vanish. The authors should report the A2A link budget, validate the slot-time feasibility, and include a sensitivity analysis with respect to synchronizatio","section":"Section 3.2, Eqs. (7) and (8)"},{"comment":"Problem P1 in Section 4 optimizes the binary communication schedule Phi, but the MDP simplification with DPAM removes the beta_{i,j}(t) decision from the action space and instead sets it deterministically by whether SNs are within the AAV communication radius. This is a heuristic, not a learned scheduling policy. The paper therefore does not solve P1 as stated: only trajectories are optimized, and the schedule is an input rule. The authors should either reformulate P1 to include the DPAM mapping as a constraint, or characterize the suboptimality of the proximity-based schedule. This is load-bearing for the 'joint trajectory and communication scheduling' contribution claimed in the Introduction and Section 4.","section":"Section 5.1, DPAM vs. Section 4.1"},{"comment":"The claimed trend that the time-average AoI decreases as the number of SNs increases is counterintuitive and is not explained by the model. With fixed AAV resources and total data volume, adding SNs should generally increase per-SN waiting time. The paper's explanation that 'more SNs allow the system to gather and transmit more frequent updates' does not follow from Eq. (10). Additionally, Figs. 4-8 show single-run curves without error bars or seed statistics, so the convergence and comparison claims are not statistically grounded. Please provide multiple-seed means/confidence intervals and re-examine the mechanism behind Fig. 6.","section":"Section 6.3, Figs. 5 and 6"}],"minor_comments":[{"comment":"Repeated grammar issues: 'a AAV' should be 'an AAV'; 'characteristize' in Section 3.2 should be 'characterize'; Table 2 has 'Y min, X max' in the Y-border row.","section":"Throughout"},{"comment":"The sentence 'the actor network combines LNGRU and SE block to generate action at based on the state sequence' appears twice in the same paragraph, once before the detailed description; please remove the duplicate.","section":"Section 5.3"},{"comment":"Eq. (23) defines the temperature alpha as an average log-probability, which is not the standard SAC temperature update. In SAC, alpha is learned by minimizing a different objective. Please clarify or remove this equation, as it is not used in the algorithm.","section":"Section 5.2, Eq. (23)"},{"comment":"The SE block uses H and W for the feature map, but the LNGRU hidden state h_t is a vector, not a 2D feature map. Please define the dimensions of h_t and how global average pooling is applied in this context.","section":"Section 5.3, Eq. (28)"},{"comment":"The quantities K and N_b are used in the training-phase complexity expression before being defined. Please define them explicitly, including the condition for when the replay buffer triggers updates.","section":"Section 5.4"},{"comment":"Several parameters needed to reproduce the simulation are missing: transmit powers P_i and P_j', channel bandwidths B and B_j', reference channel gain rho0, path-loss exponent alpha, and sensor data sizes D_i. A2A parameters are especially important given the concerns in Major Comment 2.","section":"Section 6.1, Table 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a straightforward extension of the authors' ICC 2025 paper and does not provide code or data, which weakens reproducibility. The main blocking issue is the combination of a non-standard AoI metric and unverified A2A/beamforming idealizations: these are not merely presentational, as they directly generate the reported numerical gains. A revision that replaces or justifies Eq. (10), verifies the time-budget/phase-alignment assumptions, and reformulates the DPAM scheduling rule would make the claims credible. The pruning and ablation experiments are interesting but do not compensate for these modeling issues."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a competent, incremental systems paper in the UAV-relay + DRL space. The genuinely new bit is the joint formulation of multi-AAV distributed beamforming with an AoI objective, plus the SAC-TLS variant (temporal sequence input, LNGRU, and SE block). None of the components are new, but the combination is, and the paper states it honestly.\n\nWhat the paper does well: the system model is unusually complete for this genre — G2A, A2A, and A2G phases are all written down, the energy model is standard, and the problem formulation is plausible. The related work is extensive and the comparison table is useful. The ablation studies are a good idea and the pruning analysis adds a practical angle. Simulated gains (17.3% AoI reduction, 24.5% reward, 1.6x convergence) are consistent with the model.\n\nThe soft spots, in rough order of severity:\n\n1. The A2A broadcast assumption. Eq. (7) defines δA2A(t), but the text then assumes all data is reliably broadcast within the slot. The authors never check that δG2A+δA2A+δA2G+δmove ≤ 1 for the trained policy, and they don't report the A2A parameters (B_j', P_j', ρ0, α, D_i). If δA2A is non-negligible, the A2G phase shrinks, Q(t) drops, and AoI rises for every algorithm. The comparison may still hold, but the claimed margins could shrink. This is not a fatal flaw — it's an unverified parameter regime — but it should be checked or stated as a limitation.\n\n2. The AoI update in Eq. (10) is non-standard. Multiplying the age by (1-Q(t)) when a sensor is served is not the usual reset/linear dynamics. It needs a citation or a derivation. This is the core performance metric, so it deserves more care.\n\n3. Reward design is aligned with the evaluation metrics (AoI and energy), which is normal for RL, but it means the reported outperformance is partly self-referential. That's true of most RL papers, so only a minor caution.\n\nReproducibility is limited: no code, no error bars. But the simulation logic is internally coherent.\n\nBottom line: this paper is for readers working on AoI-aware UAV/AAV networks and DRL. It deserves a serious referee — I'd send it to review — but the authors need to verify the A2A timing and justify the AoI model. The central idea holds up, the soft spots are addressable.","headline":"A competent, incremental DRL-for-UAV-relay paper; the A2A broadcast assumption is the main unverified lever, but the work is deserving of a serious referee.","tokens_in":22124,"tokens_out":4423,"would_cite":false,"duration_ms":43715,"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":"The paper claims that AAV swarms using distributed beamforming can cut IoT data age without return flights, and that a modified soft actor-critic algorithm (SAC-TLS) jointly learning trajectories and schedules achieves this in simulation.","keywords":["age of information","distributed beamforming","virtual antenna array","AAV relay","deep reinforcement learning","soft actor-critic","trajectory optimization","IoT data forwarding"],"falsifier":"A simulation or field test that replaces the lossless air-to-air broadcast with a finite-rate, lossy channel, or that adds random oscillator phase offsets to the beamforming sum in Eq. (8), would settle it: if the average AoI no longer falls by the reported 17.3%, the gain depends on those idealizations.","tokens_in":21213,"feed_emoji":"📡","tokens_out":5281,"duration_ms":58450,"temperature":0.7,"pith_summary":"This paper tries to establish that a swarm of single-antenna autonomous aerial vehicles can relay time-sensitive IoT data to a distant base station by acting as a virtual antenna array, instead of flying back and forth between sensors and the base station. If true, the propulsion-repositioning cycle that dominates age of information in drone-relay networks would be replaced by a low-latency, energy-efficient beamforming link. The paper further claims that the resulting joint trajectory-and-scheduling problem, though non-convex and time-varying, can be solved online by a deep reinforcement learning algorithm called SAC-TLS. Simulation evidence reports a 17.3% reduction in average AoI, a 24.5% increase in cumulative reward, and 1.6 times faster convergence relative to TD3, PPO, TQC, SAC, and greedy baselines. A sympathetic reading is that the paper's contribution is a system model plus a tailored DRL architecture that makes the joint optimization tractable in the simulated scenarios.","feed_headline":"Swarm beamforming cuts IoT data age by 17.3%","feed_subtitle":"Drones relay sensors as one virtual antenna; a DRL scheduler plans their paths to keep information fresh.","key_machinery":"The virtual antenna array (VAA) is the central physical mechanism: AAVs jointly transmit to the base station with an SNR given by the squared sum of per-AAV amplitude terms, (sum over j of sqrt(P_j(t) g0 d_{j,BS}^{-alpha}))^2 / sigma^2, which defines the air-to-ground rate in Eq. (8). This coherent beamforming gain is what removes the need for return flights. The second mechanism is the SAC-TLS algorithm, which stacks temporal sequence input, layer-normalized GRU gates, and squeeze-and-excitation channel attention on top of soft actor-critic, and uses dynamic proximity-based action mapping (DPAM) to replace risky binary communication decisions with deterministic proximity-based ones.","core_discovery":"The central claim is that age of information in AAV-relayed IoT networks can be minimized by having AAVs collectively form a virtual antenna array for the air-to-ground hop, so that sensors' data reaches a remote base station without requiring the AAVs to physically return to it. The paper formulates a non-convex, mixed-integer optimization that jointly chooses AAV hover trajectories and sensor communication schedules to minimize time-averaged AoI and AAV energy consumption. To solve it, the paper proposes SAC-TLS, which augments soft actor-critic with temporal sequence state input, a layer-normalized gated recurrent unit to capture long-term dependencies, and a squeeze-and-excitation block","pith_inferences":["If the lossless air-to-air broadcast assumption is relaxed, the time-slot budget must explicitly allocate air-to-air transmission, and the policy would need to trade collection breadth against intra-swarm delivery; the current DPAM scheduler would likely need an extra air-to-air-aware reward term.","A hardware consequence of the coherent-beamforming model is that oscillator synchronization and positioning accuracy become first-order constraints; sub-wavelength phase errors would degrade the virtual-antenna gain and convert the reported gains into ordinary multi-hop diversity gains.","The same SAC-TLS architecture (temporal sequence input, recurrent normalization, channel attention) could apply to other time-critical mobile relay settings, such as ground robots or mixed aerial-terrestrial relays, whenever age of information is the metric.","A testable extension is to replace the fixed SNR model with measured channel data and compare SAC-TLS's learned trajectories against a lower-bound oracle that knows future data arrivals, isolating how much of the gain comes from prediction versus exploration."],"forward_implications":["If the system works as simulated, a swarm of single-antenna AAVs can serve as a long-range relay to a distant base station without periodic return flights, eliminating the main source of AoI spikes in drone-relayed IoT.","Jointly optimizing trajectories and communication schedules in one DRL policy outperforms treating them separately; the ablation results indicate that temporal sequence input, LNGRU, and the SE block each contribute to the improvement.","The learned policy scales with swarm and network size: more AAVs reduce average AoI at the cost of higher energy, more sensors reduce average AoI, and SAC-TLS keeps the best trade-off in each tested configuration.","SAC-TLS retains near-original performance after 90% structured pruning (about 20% reward loss under tuned regularization), suggesting the trained policy can run on resource-constrained onboard hardware.","Faster convergence and higher cumulative reward imply the method operates as an online controller rather than an offline planner, useful in environments with changing data arrivals and channel conditions."],"supporting_citations":[{"why":"supplies the virtual-antenna-array SNR model whose squared-sum amplitude form defines the air-to-ground rate in Eq. (8).","marker":"[48]"},{"why":"supplies the soft actor-critic algorithm that SAC-TLS extends with temporal sequence input, LNGRU, and SE blocks.","marker":"[52]"},{"why":"supplies layer normalization used to stabilize the GRU gates in LNGRU.","marker":"[53]"},{"why":"supplies the squeeze-and-excitation block used to recalibrate channel-wise features in the actor network.","marker":"[54]"},{"why":"supplies the dynamic proximity-based action mapping (DPAM) that turns binary communication decisions into deterministic proximity-based actions.","marker":"[51]"},{"why":"provides the TD3 baseline used in performance comparisons.","marker":"[55]"},{"why":"provides the PPO baseline used in performance comparisons.","marker":"[56]"},{"why":"provides the TQC baseline used in performance comparisons.","marker":"[57]"},{"why":"supplies the dynamic structured pruning method used in the compression experiments showing 90% pruning with about 20% reward loss.","marker":"[59]"}],"fun_headline_variants":["Drone swarm beamforming cuts IoT data age","AAV cluster beams data to lower info age and energy","Multi-AAV beamforming minimizes AoI for IoT","DRL-driven drone beamforming freshens IoT data","Distributed beamforming by AAVs reduces IoT AoI"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The system assumes every AAV can reliably broadcast all collected data to every other AAV within each time slot and that distributed beamforming achieves perfect phase alignment, so the only bottlenecks are the ground-to-air collection and the air-to-ground link.","fun_headline_variants_meta":{"raw":{"variants":["Drone swarm beamforming cuts IoT data age","AAV cluster beams data to lower info age and energy","Multi-AAV beamforming minimizes AoI for IoT","DRL-driven drone beamforming freshens IoT data","Distributed beamforming by AAVs reduces IoT AoI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000334,"raw_usage":{"total_tokens":1723,"prompt_tokens":807,"completion_tokens":916,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":837}},"tokens_in":551,"tokens_out":916,"duration_ms":10442,"temperature":1.0,"reasoning_tokens":837,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:32:16.555158+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A simulation or field test that replaces the lossless air-to-air broadcast with a finite-rate, lossy channel, or that adds random oscillator phase offsets to the beamforming sum in Eq. (8), would settle it: if the average AoI no longer falls by the reported 17.3%, the gain depends on those idealizations.","supporting_citations":[{"cited_title":"Collaborative ground-space communications via evolutionary multi-objective deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"supplies the virtual-antenna-array SNR model whose squared-sum amplitude form defines the air-to-ground rate in Eq. (8)."},{"cited_title":"Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor,","cited_arxiv_id":null,"evidence_quote":"supplies the soft actor-critic algorithm that SAC-TLS extends with temporal sequence input, LNGRU, and SE blocks."},{"cited_title":"Layer normalization,","cited_arxiv_id":null,"evidence_quote":"supplies layer normalization used to stabilize the GRU gates in LNGRU."},{"cited_title":"Squeeze-and-excitation networks,","cited_arxiv_id":null,"evidence_quote":"supplies the squeeze-and-excitation block used to recalibrate channel-wise features in the actor network."},{"cited_title":"Aoi optimization for uav-assisted wireless sensor networks,","cited_arxiv_id":null,"evidence_quote":"supplies the dynamic proximity-based action mapping (DPAM) that turns binary communication decisions into deterministic proximity-based actions."},{"cited_title":"Addressing function approximation error in actor-critic methods,","cited_arxiv_id":null,"evidence_quote":"provides the TD3 baseline used in performance comparisons."},{"cited_title":"Con- trolling overestimation bias with truncated mixture of continuous distributional quantile critics,","cited_arxiv_id":null,"evidence_quote":"provides the TQC baseline used in performance comparisons."},{"cited_title":"Compressing deep reinforcement learning networks with a dynamic structured pruning method for autonomous driving,","cited_arxiv_id":null,"evidence_quote":"supplies the dynamic structured pruning method used in the compression experiments showing 90% pruning with about 20% reward loss."}],"review_version":1}