{"id":"ab4993b2-fc16-4bff-a042-1f2b7d2f4259","arxiv_id":"2505.08419","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"HMR-ODTA, an auction-based scheduler with temporal-network-based rescheduling, is claimed to reduce penalties by about 50% to 63% over prior methods in simulated hospital deliveries.","lead":"This paper introduces HMR-ODTA, a decentralized auction algorithm for scheduling online delivery tasks among heterogeneous hospital robots. In simulation, it reports roughly 50% to 63% lower late-delivery penalties than two older methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Energy model in Eqs. 3-5 is dimensionally inconsistent and its magnitude is never validated, so STN feasibility and charging decisions are unsupported; the claimed 50-63% penalty reduction lacks a physically grounded anchor.","rationale":"The reader and I identify the same load-bearing fault: the physical validity of the energy-consumption model underpinning the STN feasibility checks. I independently verified the dimensional inconsistency in Eq. 3 and Eq. 4 and traced its role into Section 5.2's STN time-window construction, which decides whether a task can be inserted and whether a charging detour is required. Since every acceptance and penalty in the simulation flows through that gate, a wrong energy equation can create or destroy the reported gap. The reader also flags the incorrect time-complexity justification (Section 5.8 claims len(SRL) is constant due to Eq. 7, but Eq. 7 bounds only per-request demand, not the count of requests) and the lack of error bars or statistical detail; those are real but secondary. No code is released, the baselines are re-implemented by the authors, and no internal inconsistency in the auction logic itself was found that would independently doom the algorithm. However, the energy model flaw alone is sufficient to reject the quantitative claim as unsupported, so the verdict remains REJECT. I did not find evidence of fraud or intent; the critique is about the argument's grounding.","tokens_in":22548,"tokens_out":2298,"duration_ms":20418,"concrete_test":"Re-run the Scenario 1 and 2 experiments with Eq. 4 replaced by a dimensionally consistent energy model calibrated against a published wheeled-robot energy curve, e.g., E_travel = P(v)*t and E_payload = c_payload * payload * distance, while keeping the STN auction logic unchanged. If the penalty gap versus EEPI/GA-MR, averaged over the four deadline configurations, shrinks by more than half in any configuration or the ranking reverses, the headline reductions are artifacts of the defective equation rather than of HMR-ODTA. Also report rejected-request counts and penalties for a variant with energy constraints disabled.","verdict_should_be":"REJECT","load_bearing_attack":"The central performance claim is a comparison against re-implemented baselines in simulation, and the simulator's task acceptance and penalty numbers are governed by the energy-consumption equations in Section 4.4. Eq. 3 adds (1/2)*v_C^2*W_RC, which has units of kg m^2/s^2 (energy), while the second term mu*g*v_C*W_RC*t1 has units kg m^2/s^2*s = kg m^2/s (not energy); the two terms cannot be summed. Eq. 4 multiplies the whole bracket (E1 + 1/2 v^2 + mu*g*v_m*t) by (W_RC + sum of payload masses), treating an energy and an acceleration-time product as dimensionless multipliers, yielding kg^2 m^2/s^2-like quantities rather than joules. Eq. 5 multiplies power by min(distance/velocity), which is time, so it is dimensionally power*time and is inconsistent with the preceding equations' scaling by mass. If these quantities are used to prune schedules and trigger charging detours as claimed in the STN interval construction of Section 5.2, then both the feasibility of accepted tasks and the comparison of penalties are not physically grounded. This is not merely an aesthetic units issue: different baselines likely respond differently to infeasible charging windows, so the 63%/50% penalty reductions could be artifacts of the simulator's flawed energy gating rather than of the allocation algorithm. The manuscript provides no calibration, no ablation without the energy model, and no sensitivity analysis to the friction coefficient or payload-weight factors, so the claim is not independently supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HMR-ODTA, a decentralized auction-based online task allocation algorithm for a team of heterogeneous mobile robots solving a multi-pickup and delivery problem with time windows (MPDPTW) in a hospital-like environment. Each robot maintains a service request list and uses a Simple Temporal Network (STN) to check feasibility and compute bids that combine penalty, efficiency, remaining energy, and completion time. New service requests are auctioned among robots, and the winning robot inserts the request into its schedule with dynamic rescheduling. The paper reports ROS/Gazebo simulations comparing HMR-ODTA with re-implemented baselines EEPI and GA-MR, claiming about 63% reduction in penalties for 40-160 tasks and about 50% for 160-280 tasks, along with lower rejection counts.","tokens_in":22958,"tokens_out":8190,"duration_ms":81197,"significance":"The problem is relevant and timely: online, heterogeneous, multi-robot pickup-and-delivery with hard and soft deadlines is an active area, and a decentralized auction with STN-based schedule feasibility is a reasonable architectural contribution. The authors also report 50 simulation runs per configuration and a non-parametric significance test, which is a good-faith attempt at statistical rigor. If the central claims are validated, the paper would provide a useful benchmark and algorithm for hospital/logistics multi-robot systems. However, the current validation is not sound: the energy-consumption equations used to gate schedule feasibility are dimensionally inconsistent, the complexity proof is incorrect, the tie-breaking rule in the auction is ambiguous, and the empirical comparison is not independently reproducible because the baselines are reimplemented and no code or data are provided.","major_comments":[{"comment":"Eq. (4) is dimensionally inconsistent. The bracketed expression adds E1 (an energy, kg m^2/s^2), 1/2 (v_RC)^2 (m^2/s^2, no mass factor), and mu*g*v_m*t (m^2/s^2, no mass factor), and then multiplies the entire bracket by (W_RC + sum of payload masses) in kg. The resulting quantity is not an energy, and the friction coefficient also changes from mu=0.4 in Eq. (3) to mu=0.02 in Eq. (4) without comment. Since Section 5.2 uses energy considerations to decide when a charging detour is needed, and Algorithm 3 computes UsedEnergy and energyRem from these quantities, the STN feasibility checks and auction bids are not physically grounded. The authors need to either replace this with a dimensionally correct energy model, or explicitly demonstrate that all reported results are insensitive to the energy model; the current simulation evidence does not support the claimed penalty reductions.","section":"Section 4.4, Eq. (4)"},{"comment":"The time-complexity proof is flawed. The claim that the number of STN time nodes, 2*len(SRL), is 'constant due to constraints 7' is incorrect: Eq. (7) bounds each individual demand, not the number of tasks in an SRL, and the capacity constraint in Eq. (8) limits total payload, which still allows arbitrarily many low-demand tasks. Therefore s grows with the input, the O(s^3) Floyd-Warshall cost inside Algorithm 3 cannot be treated as O(C), and the stated overall complexity O(|J|mn) is not established. A correct bound must account for the growth of |SRL|, or the STN size must be bounded by an actual algorithmic constant.","section":"Section 5.8"},{"comment":"The tie-breaking rule in the auction is described inconsistently between text and pseudocode. The text says that when penalties are equal the task should be assigned to the robot with lower eta, but the condition 'eta > new etaC' in Algorithm 4 selects a higher eta, and a later bullet similarly says 'Opting for a bid with lower remaining energy' while the code condition 'energyRem > new energyRem' selects higher remaining energy. Because tie-breaking determines which robot wins a bid and therefore affects the final assignment, the implemented behavior needs to be stated unambiguously and matched to the pseudocode.","section":"Section 5.6 / Algorithm 4"},{"comment":"The empirical comparison is reported only as means. Figures 3 and 4 plot averages over 50 runs without error bars, confidence intervals, or any dispersion measure, even though 50 runs per configuration are available. Section 7.5 mentions a two-tailed Wilcoxon signed-rank test but gives no p-values, no test statistics, and no table of results; the one non-significant case (UR-UA, DD=2E, penalty) is mentioned but not analyzed. Additionally, both baselines were re-implemented by the authors with no external code, data, or baseline benchmark results, so the comparison cannot be independently checked. The paper should show the full distributions, complete significance-test results, and make the implementations and data available.","section":"Section 7 / Figures 3-4"},{"comment":"The abstract's headline percentages ('nearly 63%' for 40-160 tasks, 'approximately 50%' for 160-280 tasks) are not traceable to a single comparison in Tables 5-6. Depending on the deadline configuration and scenario, the reported EEPI-vs-HMR-ODTA penalty reduction for 160-280 tasks ranges from 32.94% to 85.04%, and the corresponding rejection reduction ranges widely as well. The authors should define exactly which aggregation (e.g., across all deadline settings and both scenarios) produces the abstract numbers, and report the associated variance.","section":"Abstract / Tables 5-6"}],"minor_comments":[{"comment":"The text says 'octal distance between map points'; this should presumably be 'geodesic distance' or 'Euclidean distance', since the distances are computed with jump point search.","section":"Section 6.1"},{"comment":"The sentence 'GA-MR shows an improvement of 15.91% over GA-MR' should read 'EEPI shows an improvement of 15.91% over GA-MR'; the current wording is self-referential and confusing.","section":"Section 7.3"},{"comment":"The pseudocode assigns currPos=node before popping the next node from the schedule, so currPos receives the just-completed node rather than the next destination. The variable 'update' in Line 12 is also undefined. Please correct the ordering and define the intended update condition.","section":"Algorithm 5"},{"comment":"The notation R_X is used both for a class index and as a set of robots, and the returned robot R_Y^X is not formally introduced. Please disambiguate the notation.","section":"Algorithm 2"},{"comment":"The GA-MR baseline is cited as [16] in the contributions list but as [14] in Section 2 and Section 6; this inconsistency should be fixed.","section":"Introduction / Related Work"}],"recommendation":"major_revision","confidential_remarks":"The central mechanism is plausible and the topic fits the journal, but the current experimental evidence is not reliable: the energy model in Eq. (4) is dimensionally invalid, the complexity proof is wrong, and the baseline comparison is not externally reproducible. These are fixable with a substantial revision, but not with cosmetic changes. I would advise the editor to ask for corrected equations, rerun or re-justified simulations, complete statistical reporting, and a clear statement of the implemented tie-breaking rules before considering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick context: the paper combines auction-based task allocation with Simple Temporal Network (STN) scheduling for online heterogeneous pickup-and-delivery with time windows, plus rescheduling. That combination is genuinely new in the MAPDP literature, and the authors identify a real gap: most prior work is offline, homogeneous, or lacks dynamic rescheduling. They also ran a substantial simulation campaign: 50 trials, four deadline configurations, two robot distributions, and they report Wilcoxon tests. The algorithm description is mostly clear. I'd believe the auction mechanism and the STN rescheduling idea are worth exploring.\n\nThe problem is the energy model and its role in the experiments. Equations 3 and 4 are dimensionally inconsistent: Eq. 3 adds a kinetic-energy term (kg·m²/s²) to a friction term that has units kg·m²/s, and Eq. 4 multiplies an entire energy bracket by payload mass, producing units like kg²·m²/s². Eq. 5 is power times time, which is fine, but it doesn't match the mass-scaled form of the others. These equations are not decorative. They determine whether a robot has enough energy to accept a task, when it must detour to charge, and whether a schedule is feasible. If the numbers are unphysical, the simulator's accept/reject behavior is ungrounded, and the claimed 63%/50% penalty reductions could be artifacts of the energy gate rather than of the auction logic. The paper gives no calibration, no sensitivity analysis, and no ablation without the energy model, so the central empirical claim is not independently supported.\n\nThere are also smaller issues. The time-complexity proof in Section 5.8 is wrong: it treats a bounded SRL size as a constant without stating it as an assumption, and the O(len(SRL)) claim in Algorithm 5 is hand-wavy. Figures 3 and 4 show averages without error bars or variance, which is surprising after 50 runs. Both baselines are author re-implementations with no released code, which is common but increases the risk of inadvertent bias, especially given the energy model issues.\n\nThis is not a case of circularity in the algorithm itself. The novelty is real, and the authors are honest about their limitations. But the load-bearing empirical evidence is not currently credible. The fix is achievable: correct and validate the energy equations (or drop them), add variance and sensitivity analysis, fix the complexity proof, and release code.\n\nI'd send this to peer review because the problem is important and the proposed algorithm deserves scrutiny from the community. But I would not accept it as-is, and I wouldn't cite its performance numbers until the energy model is fixed and the comparisons are reproduced. Worth a serious referee, not a desk rejection.","headline":"A genuinely new online heterogeneous pickup-and-delivery auction algorithm, but the energy model is dimensionally broken and the performance claim rests on it; needs major revision before it can be trusted.","tokens_in":23427,"tokens_out":4203,"would_cite":false,"duration_ms":38913,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"HMR-ODTA claims that a decentralized auction with local re-scheduling reduces late-delivery penalties by about 63% on smaller task sets and about 50% on larger ones, relative to two reimplemented methods, in simulated hospital settings.","keywords":["Multi-robot task allocation","Pickup and delivery problem","Time windows","Heterogeneous robots","Online scheduling","Auction-based allocation","Simple Temporal Network","Penalty minimization"],"falsifier":"Re-run the same experiments with robot energy consumption measured on physical hardware (or a validated electro-mechanical model) instead of the Section 4.4 equations; if the penalty and rejection gaps over EEPI and GA-MR shrink below the reported 50–63% and 26–51% levels, the scheduling advantage is an artifact of the energy model rather than of the auction-and-rescheduling mechanism.","tokens_in":22392,"feed_emoji":"🤖","tokens_out":5629,"duration_ms":56915,"temperature":0.7,"pith_summary":"Coordinating delivery requests that arrive over time—each with a pickup point, a drop-off point, and a hard or soft deadline—is hard when the robot team is heterogeneous. This paper proposes HMR-ODTA, a decentralized auction in which every robot bids on each new request, and the bid comes from a Simple Temporal Network check of whether and where the request fits into that robot's existing schedule. The paper reports that, in 50 simulated hospital runs per setting, HMR-ODTA reduces cumulative late-delivery penalties by nearly 63% for 40–160 requests and roughly 50% for 160–280 requests compared with two reimplemented methods, EEPI and GA-MR, and also rejects fewer requests. The reason matters: fixed assignments fail when new urgent requests appear, while letting each robot re-insert requests into its own schedule keeps the team responsive without a central planner.","feed_headline":"Auction-based scheduler cuts hospital delivery penalties by up to 63%","feed_subtitle":"HMR-ODTA lets each robot re-plan its own deliveries in real time, beating EEPI and GA-MR in a simulated hospital.","key_machinery":"The Simple Temporal Network (STN)—a directed graph whose edges encode allowable time differences between actions—is the load-bearing object. Each robot builds an STN over its accepted service requests and asks, for every new request, whether the request can be inserted at some position without violating deadlines, capacity, and energy limits, including detours to a charging station. The bid a robot submits is a tuple (penalty, efficiency η, remaining energy), compared lexicographically: lowest penalty wins, ties go to the least efficient robot, and further ties to the robot with lower remaining energy, reserving more capable robots for future critical requests. A round-robin modulo rule picks which robot runs each auction, keeping the system decentralized.","core_discovery":"The central claim is that an online, decentralized auction—where each robot bids using a Simple Temporal Network feasibility check plus a composite score of penalty, efficiency, and remaining energy—can coordinate a heterogeneous robot fleet for pickup-and-delivery with time windows, and that this reduces cumulative late-delivery penalties by roughly 63% for 40–160 tasks and 50% for 160–280 tasks compared with reimplemented EEPI and GA-MR baselines. The paper also reports fewer rejected service requests at every tested load. The mechanism credited for the improvement is rescheduling: unlike the baselines, which keep assignments fixed once made, HMR-ODTA lets a robot re-insert its accepted requests into its local schedule whenever a new request arrives, so delayed or newly arriving tasks can be accommodated without rebuilding a global plan.","pith_inferences":["The same auction-plus-STN mechanism could be extended to robot failures: a robot that becomes unavailable mid-shift could re-auction its accepted requests, which the paper leaves to future work.","Since the bid is only a local insertion check, an open question is how far HMR-ODTA sits from an offline optimal schedule; a competitive-ratio bound would require a formal analysis the paper does not provide.","The reported gains may be sensitive to the energy-consumption formulas in Section 4.4, which mix kinetic and frictional terms; correcting those equations and re-running the comparison would reveal whether the 50–63% penalty reduction is a scheduling effect or an artifact of the energy model."],"forward_implications":["In the simulated hospital environment, HMR-ODTA completes more requests within deadlines and incurs lower cumulative lateness than EEPI and GA-MR at every tested request volume from 40 to 280.","The penalty reduction holds for both equal and unequal robot distributions and across deadline configurations E, 2E, [5E, 10E], and [E, 10E], with one reported non-significant case: average penalty for unequal robots with unique attributes when the deadline is 2E.","HMR-ODTA cuts the number of rejected hard-deadline requests as well as penalties, because a request that cannot fit one robot's current schedule may still fit another robot's schedule after that robot re-inserts it.","The per-request auction is computationally light, with each robot's bid computed in constant time in practice and overall complexity O(|J|mn) for |J| requests, m robot classes, and n robots per class.","Because the scheduler explicitly checks charging detours, the STN inserts not just a pickup and drop-off but also a possible charging stop, which is how the algorithm keeps robots from becoming stranded mid-route."],"supporting_citations":[{"why":"Defines the Simple Temporal Network formalism used for schedule feasibility and request insertion.","marker":"[12]"},{"why":"EEPI baseline re-implemented for comparison on rejected requests and penalties.","marker":"[13]"},{"why":"GA-MR baseline re-implemented for comparison on rejected requests and penalties.","marker":"[14]"},{"why":"Improved jump point search used to precompute geodesic distances between map nodes, feeding the travel-time and energy estimates.","marker":"[41]"},{"why":"Supplies the method for generating real-time deadline configurations used in the experiments.","marker":"[47]"},{"why":"Provides the hospital floor plan used to build the simulation environment.","marker":"[38]"}],"fun_headline_variants":["Decentralized auction cuts robot delivery penalties by 63%","Online auction scheduler reduces late deliveries by 63%","Robot team auction cuts delivery penalties up to 63%","HMR-ODTA rescheduling cuts penalties by 50-63%","Auction-based rescheduling cuts hospital delivery penalties 63%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed advantage rests on the energy-consumption formulas in Section 4.4 being physically accurate, because those formulas decide which routes and charging detours are feasible; if they do not reflect real robot energy use, the schedules and the simulated comparison would not carry over to practice.","fun_headline_variants_meta":{"raw":{"variants":["Decentralized auction cuts robot delivery penalties by 63%","Online auction scheduler reduces late deliveries by 63%","Robot team auction cuts delivery penalties up to 63%","HMR-ODTA rescheduling cuts penalties by 50-63%","Auction-based rescheduling cuts hospital delivery penalties 63%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000801,"raw_usage":{"total_tokens":3519,"prompt_tokens":942,"completion_tokens":2577,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":2491}},"tokens_in":558,"tokens_out":2577,"duration_ms":16791,"temperature":1.0,"reasoning_tokens":2491,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:55:21.800983+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same experiments with robot energy consumption measured on physical hardware (or a validated electro-mechanical model) instead of the Section 4.4 equations; if the penalty and rejection gaps over EEPI and GA-MR shrink below the reported 50–63% and 26–51% levels, the scheduling advantage is an artifact of the energy model rather than of the auction-and-rescheduling mechanism.","supporting_citations":[{"cited_title":"Intelligent Service Robotics14(5), 707–727 (2021) https: //doi.org/10.1007/s11370-021-00393-4","cited_arxiv_id":null,"evidence_quote":"GA-MR baseline re-implemented for comparison on rejected requests and penalties."},{"cited_title":"Algorithms15(6), 198 (2022) https: //doi.org/10.3390/a15060198","cited_arxiv_id":null,"evidence_quote":"Improved jump point search used to precompute geodesic distances between map nodes, feeding the travel-time and energy estimates."},{"cited_title":"partitioned fixed- priority and edf scheduling for hard real time","cited_arxiv_id":null,"evidence_quote":"Supplies the method for generating real-time deadline configurations used in the experiments."},{"cited_title":"https://github.com/aws-robotics/ aws-robomaker-hospital-world Accessed 2023-05-24","cited_arxiv_id":null,"evidence_quote":"Provides the hospital floor plan used to build the simulation environment."}],"review_version":1}