{"id":"16772817-defa-4dc3-9513-5778d2c3ce16","arxiv_id":"2505.06997","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A learning-based controller for coordinating humans, drones, and ground vehicles in emergency sensing claims an average 18.42% increase in task completion rate over baseline methods.","lead":"This paper presents a learning-based method that coordinates humans, drones, and ground robots to complete sensing tasks during emergency rescue. In simulations and a real-world district, it reports higher task completion rates than four baseline methods, with an average improvement of 18.42%.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 18.42% average TCR improvement matches only the training-performance column of Table II against one baseline; the full reported comparisons average about 11%, and two Table II cells contradict the claim that HECTA4ER consistently outperforms MANF-RL-RP.","rationale":"The reader's CONDITIONAL verdict is appropriate. The reader's formal 'weakest_assumption' field points to model simplification, which is a real limitation, but the single most load-bearing issue for the central claim is the numerical support for the headline result: the 18.42% advantage appears to be computed from a subset of the available comparisons, and two cells in Table II contradict the claim of consistent superiority. This is not a matter of theoretical disagreement with a baseline; it is an internal check on whether the paper's own data support its abstract. The convergence theorem in Appendix D is also not a valid finite-step convergence proof, but that affects the theoretical framing rather than the empirical central claim. The available tables still show HECTA4ER ahead of MANF-RL-RP in most cells, so the result is plausibly positive but overstated as written. Thus the reader's CONDITIONAL verdict should stand, with the condition that the headline statistic be recomputed and scoped, the inconsistent table cells be resolved or explained, and the code and seeds be released for audit.","tokens_in":24773,"tokens_out":7820,"duration_ms":72378,"concrete_test":"Audit the headline statistic by recomputing the average TCR improvement over every reported comparison: all five columns of Table II, Table IV, and the per-seed data underlying Figure 9, and state the exact set of comparisons and baselines included; if the aggregate average is below 18.42%, or if rows 6-1 and 8-2 reverse under the released experimental seeds, revise the abstract and Section VI accordingly. Release the random seeds, environment configurations, and code so that the two inconsistent Table II cells can be independently reproduced.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and conclusion state that HECTA4ER achieves an average 18.42% increase in TCR over baseline methods. Recomputing from Table II, the mean difference in the 'Training performance' column across scenarios 5-1 through 10-3 is about 18.5%, which matches the headline, but this column covers only base-scenario performance against one baseline, MANF-RL-RP. Averaging all five robustness columns in Table II (85 comparisons) gives about 11.1%; the case-study Table IV advantage over MANF-RL-RP averages about 9.9%; and Figure 9, which includes other scenarios and baselines, is not part of the stated denominator. The text also says HECTA4ER 'consistently outperforms' MANF-RL-RP, yet rows 6-1 and 8-2 in the 'Changing sensing entity position' column show MANF-RL-RP higher (43.2 vs 43.0 and 30.5 vs 30.4, respectively). Because no code, per-seed results, or trained models are released, the reader cannot audit how 18.42% was derived or whether those two negative cells are typos. The central empirical claim is therefore not supported by the paper's own tables in the form stated, and the headline quantitative result needs correction or a precisely scoped restatement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the Heterogeneous-Entity Collaborative-Sensing Task Allocation (HECTA) problem in emergency rescue, modeling it as a Dec-POMDP over a discrete grid and proposing a MARL algorithm called HECTA4ER. The algorithm uses a Centralized Training with Decentralized Execution architecture with convolutional feature extraction, recurrent hidden states for action-observation history, and a mixing network that fuses global and local information. A key design is a 'Hard-Cooperative' policy in which UGVs interrupt their own sensing to recharge low-battery UAVs. The claims include an NP-hardness proof, a Markov property proof for belief states, a convergence theorem, extensive simulations over ten scenarios with four baselines, robustness tests, ablations, and a real-world case study, with a headline result of an average 18.42% increase in task completion rate (TCR) over baseline methods.","tokens_in":25018,"tokens_out":12745,"duration_ms":112078,"significance":"If the claims are supported, the paper offers a useful application of MARL to a realistic heterogeneous crowdsensing problem, combining partial observability, heterogeneous capabilities, energy constraints, and time limits. The most valuable parts are the explicit problem formalism, the concrete baseline comparisons (including MANF-RL-RP and FD-MAPPO), the ablation study, and the attempt to move beyond simulation with a real-world case study. The paper also states its limiting assumptions clearly. However, the central quantitative result is not supported by the tables in the form stated, and the convergence proof contains substantial gaps, so the current version needs revision before the empirical and theoretical contributions can be accepted.","major_comments":[{"comment":"The headline claim of an average 18.42% increase in TCR over 'baseline methods' is not supported by the reported tables as written. Recomputing Table II, the mean difference in the 'Training performance' column is about 18.5% across the 17 scenario rows, which can reproduce the 18.42% figure only if that single column and the single baseline MANF-RL-RP are used as the comparison set. Averaging all five columns of Table II (85 comparisons) gives about 11.1%, while the case-study advantage over MANF-RL-RP in Table IV averages about 9.9%; the other baselines (Greedy-SC-RP, FD-MAPPO, HECTA4ER-Voluntary) are not part of the claimed denominator. Furthermore, §V-B states that HECTA4ER 'consistently outperforms MANF-RL-RP in each case,' but Table II rows 6-1 and 8-2 in the 'Changing sensing entity position' column show MANF-RL-RP with higher TCR (43.2 vs. 43.0 and 30.5 vs. 30.4, respectively). The paper must either correct the headline to name the exact comparison set or provide a properly aggregated average over all baselines and scenarios, and it must reconcile or annotate the two negative cells.","section":"Abstract; §V-B, Table II; §VI"},{"comment":"The robustness protocol is under-specified. The text says that after training on a base scenario the algorithm is 'evaluated across 50 new random scenarios,' but Table II does not show 50 scenarios, per-seed results, or confidence intervals; each cell appears to be a single TCR value, and the column labeled 'Training performance' is never defined in the caption or the body. It is therefore impossible to determine whether the reported robustness results are means over 50 random scenarios, means over seeds, or single runs. The paper should state the aggregation procedure, the number of seeds, and the variance or confidence intervals for every cell in Table II.","section":"§V-B, Table II"},{"comment":"The convergence proof is not valid for the algorithm as implemented. The induction assumes at initialization that the Q-values are already within ε0 of the optimal Q* (Eq. 33), which presupposes the conclusion. The proof then applies the tabular Q-learning update rule (Eqs. 36-37), although HECTA4ER uses neural-network function approximation, target networks, an experience replay buffer, and the mixing losses of Eq. 24, none of which fit the stochastic-approximation framework invoked. Moreover, the Robbins-Monro conditions give asymptotic convergence, not 'convergence within a finite number of steps' as claimed at the end of the proof, and the inequality chain in Eqs. 43-44 requires σ<εt, which is not established. The authors should either remove the finite-step claim and supply a rigorous asymptotic result under stated assumptions, or present the convergence section as empirical and qualitative.","section":"Appendix D, Theorem 4"}],"minor_comments":[{"comment":"The 'Hard-Cooperative' override is specified only for the case where the UGV detects a UAV in its range with insufficient power; please state explicitly how the rule composes with the action-filtering mechanism of Algorithm 1 when the UGV cannot reach the UAV within one time step or the target cell is occupied by an obstacle.","section":"§III-B, Eq. 1"},{"comment":"The conditioning set in the normalization factor is written as 'I_{t-1}^t' in the denominator but should be 'I_k^{t-1}'; the same notational issue appears in Eqs. 30-31 and should be corrected.","section":"Appendix C, Eq. 28"},{"comment":"The complexity expression O(W×(B+batchsize+TimeLimit)) is not derived and appears to omit per-agent factors and environment-step costs; either provide a derivation or remove the formula.","section":"§V-A1"},{"comment":"The affiliation block lists 'Yongling Tian' while the author biography uses 'Yonglin Tian'; the authors should harmonize the spelling.","section":"Title page and biography"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of a crowdsensing or multi-agent systems venue, and its empirical effort is substantial. My main concerns are the misstated headline result, the under-specified robustness table, and the invalid convergence proof. If the authors can precisely scope the 18.42% claim, provide per-seed or interval statistics, and either repair or weaken the convergence theorem, the paper would be acceptable. I would also recommend requiring code and trained-model release as a condition of acceptance, since the central numerical claims cannot currently be audited."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid engineering paper that overstates one number. The MARL formulation for air-ground-human crowdsensing is clearly described, the experimental sweep is broad, and the real-world case study with trajectory visualizations is a genuine plus. The 'Hard-Cooperative' recharge rule is not deeply novel—UGV-carries-and-charges-UAV appears in [9] and [10]—but combining it with a Dec-POMDP treatment and action filtering is a reasonable contribution.\n\nThe soft spots are real. The abstract's headline '18.42% average TCR increase over baselines' only matches the training-performance column of Table II against MANF-RL-RP. Averaging all 85 robustness comparisons gives about 11%, and two cells (6-1 and 8-2, changing entity position) actually show MANF-RL-RP higher. That contradicts the text's 'consistently outperforms.' No code or per-seed results are released, so this cannot be audited. The claim needs to be re-scoped or the data released.\n\nSecond, the convergence theorem in Appendix D is not a proof. It assumes a Q-learning contraction with a hand-waved bound, then asserts an alpha schedule satisfies Robbins-Monro while discussing 'finite steps' in a way that is not established. Equation 44 requires sigma < epsilon_t, which is just assumed. This should be removed or replaced with a real argument, not left as a formal result.\n\nThird, the robustness table says '50 new random scenarios' but the table only gives single numbers per variation; there is no indication of variance or how the 50 scenarios aggregate. The reader flagged this and it is a fair concern. The simplifications (static 2D grid, one entity type per task) are acknowledged in the discussion, so I do not count them as hidden flaws.\n\nOverall, the empirical pattern—HECTA4ER above the baselines in most settings, the ablation showing SEDM matters, the case study showing recharge behavior—is plausible and useful. But the paper's main quantitative claim, as stated, is not supported by its own tables. That is fixable with a corrected abstract and a code/data release. I would send it to review, conditional on the authors addressing the headline number and the convergence appendix.","headline":"A useful MARL application with a broad experimental sweep, but the 18.42% headline only matches a subset of Table II, the convergence proof is not a proof, and no code is released to audit the numbers.","tokens_in":25622,"tokens_out":2503,"would_cite":false,"duration_ms":25798,"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":"This paper claims that a multi-agent reinforcement learning algorithm, HECTA4ER, using a 'Hard-Cooperative' policy where UGVs recharge low-battery UAVs, raises the average task completion rate in emergency crowdsensing by 18.42% over…","keywords":["Mobile Crowdsensing","Collaborative Sensing","Emergency Rescue","Task Allocation","Partially Observable Environmental States","Multi-Agent Reinforcement Learning","Hard-Cooperative policy","HECTA"],"falsifier":"A controlled field test in which a low-battery UAV is placed outside all UGV detection ranges and obstacles or tasks move between time steps should reproduce the 18.42% completion-rate advantage; if the advantage disappears in that setting, the Hard-Cooperative policy's guaranteed-recharge assumption is doing the work rather than the learning algorithm.","tokens_in":24537,"feed_emoji":"🚁","tokens_out":6885,"duration_ms":65898,"temperature":0.7,"pith_summary":"This paper claims that task allocation among humans, UAVs, and UGVs in emergency rescue can be modeled as a decentralized partially observable Markov decision process, and that a multi-agent reinforcement learning agent trained with centralized information can allocate tasks effectively under partial observability. The key novel mechanism is a 'Hard-Cooperative' policy in which UGVs drop their own sensing to recharge low-battery UAVs, which the authors argue is essential for sustaining UAV sensing. In simulations the proposed algorithm HECTA4ER achieves on average 18.42% higher task completion rate than baseline methods, and a real-world case study in a community suggests it maintains a higher completion rate in a dynamic setting. A sympathetic reader would care because emergency response requires decisions under limited communication and time pressure, and the paper offers an end-to-end learning approach that does not require a hand-coded planner.","feed_headline":"Cooperative rescue sensing lifts task completion by 18.42%","feed_subtitle":"Simulations and a real district test show the trained policy beats planners that ignore agent cooperation.","key_machinery":"The load-bearing mechanism is the 'Hard-Cooperative' policy: if a UGV detects that a UAV cannot move due to insufficient power, the UGV immediately goes to replace the UAV's battery instead of executing its own sensing task. This policy is expressed as a deterministic transition rule for the UGV's location. Around this policy, the algorithm HECTA4ER uses a Dec-POMDP formulation with belief states that satisfy the Markov property, a recurrent decision module that summarizes action-observation history in hidden states, an action-filtering mechanism that masks actions outside the movable range to counter sparse rewards, and a mixing network that enforces the monotonicity condition between individual and joint action values.","core_discovery":"The paper's central claim is that the Heterogeneous-Entity Collaborative-Sensing Task Allocation (HECTA) problem in emergency rescue, which is NP-hard, can be solved effectively by a multi-agent reinforcement learning algorithm that explicitly models partial observability and enforces a cooperative energy-management rule. The authors formulate the problem as a Dec-POMDP, prove that belief states based on action-observation history satisfy the Markov property, and use this to justify a recurrent decision module. They then introduce HECTA4ER, built on a Centralized Training with Decentralized Execution architecture, with a convolutional feature extractor, a recurrent policy with action filtering, and a mixing network that combines global and local information under an Individual-Global-Max consistency condition. In simulations across varied scenarios, HECTA4ER achieves an average 18.42% higher task completion rate than four baselines, and in a real-world case study in a 12 $km^{2}$ community it maintains the highest completion rate across different sensing time limits.","pith_inferences":["The paper's model ties each task to a single entity type; a natural extension would allow multi-capability tasks and ask whether the same architecture can learn synchronized multi-agent arrivals, a setting the current formulation explicitly rules out.","The failed replenishment at time T=6 in the case study (UAV outside UGV detection range) suggests a concrete algorithmic improvement the authors do not explore: letting UGVs predict UAV energy depletion before it becomes critical, rather than reacting only when the UAV is already immobile.","The convergence proof in Appendix D uses textbook Robbins-Monro conditions with a specific decreasing step size; the practical RMSprop schedule in the experiments is not shown to satisfy those conditions, so the 'finite-step convergence' is best read as an assurance argument rather than a measured guarantee."],"forward_implications":["If the 18.42% average improvement replicates, emergency responders can complete substantially more sensing tasks within a fixed time window by using the learned policy rather than greedy or independent reinforcement learning baselines.","The 'Hard-Cooperative' policy implies that UGV capacity is reserved for UAV battery support, so systems with a higher UGV-to-UAV ratio should see larger gains, while fleets with too few UGVs may leave UAVs stranded without recharging.","Because the problem is formulated as a Dec-POMDP with belief states, the same algorithm structure can be reapplied to other partially observable collaborative sensing tasks with different entity types and energy constraints.","The real-world case study suggests that even with only six sensing entities, the learned policy can outperform baselines, indicating that the approach does not require a large fleet to be useful."],"supporting_citations":[{"why":"Supplies the air-ground collaboration model in which UGVs carry and recharge UAVs, the foundation of the Hard-Cooperative policy.","marker":"[9]"},{"why":"Extends UGV-UAV recharging with hierarchical multi-agent deep RL and diffusion models; provides the movement-radius and energy-consumption settings used in the case study.","marker":"[10]"},{"why":"Provides the state-of-the-art baseline for UAV-worker-car collaborative sensing in disaster response that HECTA4ER is compared against.","marker":"[30]"},{"why":"Supplies the monotonic value-function factorization principle that underlies the mixing module and its Individual-Global-Max consistency losses.","marker":"[42]"},{"why":"Justifies the QTRAN-style consistency losses used to enforce the relationship between joint and individual action values in the mixing module.","marker":"[43]"},{"why":"Provides the fully decentralized MARL baseline (independent PPO) used in the comparisons.","marker":"[44]"},{"why":"Supplies the check-in distribution data used to model realistic initial positions and task locations in the simulations.","marker":"[45]"}],"fun_headline_variants":["Multi-agent RL for rescue sensing lifts task completion by 18.42%","Cooperative air-ground-human agents boost emergency crowdsensing success","HECTA4ER: RL coordination for rescue sensing beats planners by 18%","UAV-UGV-human teamwork in rescue: 18% higher task completion","Recharging UAVs via cooperation: new RL policy improves rescue sensing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model assumes a static 2D grid with discrete time steps, that each task can be done by exactly one entity type, and that recharging always succeeds whenever a UGV detects a low-battery UAV; if real rescue scenes involve moving obstacles, communication breakdowns, or tasks needing multiple capabilities, the central performance gain may not carry over.","fun_headline_variants_meta":{"raw":{"variants":["Multi-agent RL for rescue sensing lifts task completion by 18.42%","Cooperative air-ground-human agents boost emergency crowdsensing success","HECTA4ER: RL coordination for rescue sensing beats planners by 18%","UAV-UGV-human teamwork in rescue: 18% higher task completion","Recharging UAVs via cooperation: new RL policy improves rescue sensing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001324,"raw_usage":{"total_tokens":5435,"prompt_tokens":1039,"completion_tokens":4396,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":655,"completion_tokens_details":{"reasoning_tokens":4298}},"tokens_in":655,"tokens_out":4396,"duration_ms":30176,"temperature":1.0,"reasoning_tokens":4298,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:27:41.734146+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled field test in which a low-battery UAV is placed outside all UGV detection ranges and obstacles or tasks move between time steps should reproduce the 18.42% completion-rate advantage; if the advantage disappears in that setting, the Hard-Cooperative policy's guaranteed-recharge assumption is doing the work rather than the learning algorithm.","supporting_citations":[{"cited_title":"Air-ground spatial crowdsourcing with uav carriers by geometric graph convolutional multi-agent deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the air-ground collaboration model in which UGVs carry and recharge UAVs, the foundation of the Hard-Cooperative policy."},{"cited_title":"Energy-efficient ground-air-space vehicular crowdsensing by hierarchical multi-agent deep reinforcement learning with diffusion models,","cited_arxiv_id":null,"evidence_quote":"Extends UGV-UAV recharging with hierarchical multi-agent deep RL and diffusion models; provides the movement-radius and energy-consumption settings used in the case study."},{"cited_title":"Collab- orative route planning of uavs, workers, and cars for crowdsensing in disaster response,","cited_arxiv_id":null,"evidence_quote":"Provides the state-of-the-art baseline for UAV-worker-car collaborative sensing in disaster response that HECTA4ER is compared against."},{"cited_title":"Monotonic value function factorisation for deep multi- agent reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the monotonic value-function factorization principle that underlies the mixing module and its Individual-Global-Max consistency losses."},{"cited_title":"Qtran: Learning to factorize with transformation for cooperative multi-agent reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Justifies the QTRAN-style consistency losses used to enforce the relationship between joint and individual action values in the mixing module."},{"cited_title":"The surprising effectiveness of ppo in cooperative multi-agent games,","cited_arxiv_id":null,"evidence_quote":"Provides the fully decentralized MARL baseline (independent PPO) used in the comparisons."},{"cited_title":"Online organizing large-scale heterogeneous tasks and multi-skilled participants in mobile crowdsensing,","cited_arxiv_id":null,"evidence_quote":"Supplies the check-in distribution data used to model realistic initial positions and task locations in the simulations."}],"review_version":1}