{"id":"7c4440b9-309e-4838-bc26-5fc7da599347","arxiv_id":"2508.19731","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"HATA folds pedestrian motion patterns (Maps of Dynamics) into multi-robot task allocation via a weighted cost tuned by Bayesian optimization, reducing mission time in simulation.","lead":"A robot-fleet task allocator uses long-term records of where people walk to keep robots out of crowded paths while assigning jobs. In simulation, it finished missions up to 26% faster than a distance-only allocator and up to 19% faster than a prior human-aware baseline.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline gains may be driven by δ-threshold path avoidance rather than the stochastic allocation cost; no ablation and no variance reporting isolate HATA's contribution.","rationale":"The reader's weakest assumption concerned calibration on the same simulator and the simulator's fidelity as a proxy. That is a real concern, but there is already a train/test split: weights are fit on November 28 and evaluation is on December 12, so the calibration is not literally circular. The more load-bearing issue is construct validity: the paper's claimed novelty is the stochastic cost function in Eq. 3, yet the experimental comparison changes both the path planner and the bid cost simultaneously. With w0≈w1, the allocation cost is nearly the path length plus a small encounter penalty, while the δ-threshold in A* is a separate mechanism that can drastically alter routes and therefore mission/wait times. Table II shows δ=0.55 can be worse than HA-Alloc, so the reported 'up to 26%' is selected rather than robust. The absence of error bars or significance tests makes it impossible to tell whether the differences are real or noise. This does not refute HATA—the method may well work—but the central claim is under-supported without an ablation separating path-planning avoidance from allocation cost, and without variance reporting. The paper deserves credit for a held-out evaluation day, extensive simulation (1350 runs), and a plausible mechanism; a conditional acceptance requiring an ablation and confidence intervals is the appropriate outcome, which matches the reader's verdict. Hence no change to the verdict is needed.","tokens_in":10380,"tokens_out":5675,"duration_ms":72149,"concrete_test":"Run four configurations in the same coordination framework: (1) path-based shortest-path bids, (2) MoD-based A* with δ=0.65 but distance-only bids (w1=0), (3) full HATA, (4) HA-Alloc. For each fleet size and time window, report the distribution of mission and waiting times over the 90 seeded runs with bootstrap 95% CIs. If cell (2)≈(3), the stochastic cost function contributes nothing; if cell (1)≈(2), the reported gains are path-planning artifacts. Also report results averaged over δ∈{0.55,0.65,0.75,0.85} instead of the best δ.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline improvements compare HATA to a path-based auction baseline, but HATA changes two things at once: the path planner (A* with MoD threshold δ, Section III.D) and the bid cost (stochastic human-encounter term w1·η, Eq. 3). Since the final BO weights are w0=1.15, w1=0.95—nearly equal—the allocation cost is close to path length plus encounter probability; the larger behavioral difference may be δ-driven circumvention of high-density cells. The paper never ablates this: there is no run of MoD-avoiding A* with distance-only bids, nor a run with the HATA cost but no δ avoidance. Table II also shows that the δ sweep matters: at fleet size 15, δ=0.55 gives 281.1 s, worse than HA-Alloc's 269.9 s, so the 'up to 26%' result is a favorable point in (δ, fleet size, time-window) space, with no error bars or significance tests despite 90 runs per scenario. Moreover, the weights are calibrated in the same coordination simulator used for evaluation, so the size of the effect is only established for that simulator. Thus the central claim that the stochastic cost function improves allocation is not isolated from path-planning effects, threshold selection, or simulator-specific calibration.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HATA, a human-aware multi-robot task allocation method that augments auction-based bidding with a stochastic cost combining Euclidean path length and the probability of human encounters, where the encounter probabilities come from grid-based Maps of Dynamics (MoDs). A* path planning is modified with a threshold δ so that cells whose human-presence probability exceeds δ are circumvented. The cost weights w0 and w1 are tuned by Bayesian optimization in a coordination simulator that replays real pedestrian trajectories from the ATC dataset. Experiments with fleets of 5, 10, and 15 robots compare HATA with a path-based auction baseline and with HA-Alloc, reporting lower failure rates, mission completion times, and waiting times, with headline mission-time reductions of up to 26% over the path-based method and 19% over HA-Alloc.","tokens_in":10756,"tokens_out":4368,"duration_ms":49346,"significance":"If the reported gains are robust, the paper makes a useful contribution: it provides a concrete way to inject long-term pedestrian motion models into the bidding process of MRTA, and it evaluates on real pedestrian data with a coordination simulator, including a held-out test day. The scalability comparison against HA-Alloc is also a strength. However, the central performance claim is not yet isolated: the same simulator is used for weight calibration and evaluation, the method changes both path planning and bid cost simultaneously, and all tables report only single-point averages without variance or significance tests. I therefore regard the contribution as promising but not fully established by the current evidence.","major_comments":[{"comment":"All performance results are reported as single averages, despite 90 simulation runs per scenario. No standard deviations, confidence intervals, or significance tests are given. This matters because several reported differences are not obviously large relative to expected run-to-run variation, and Table II itself shows a non-monotone and sometimes worse HATA result (δ=0.55, 15 robots: 281.1 s vs HA-Alloc 269.9 s). The 'up to 26%' claim is therefore a selected favorable point in the (δ, fleet size, time window) space. The paper should report variance, paired comparisons over the same pedestrian replays, and a clear protocol for choosing δ without peeking at the test day.","section":"§IV.C and §V, Tables I–III"},{"comment":"The cost weights w0 and w1 are fitted by Bayesian optimization on the error between estimated and simulated mission time in the same coordination simulator that is later used to measure the reported improvements. Although November 28 is used for calibration and December 12 for evaluation, both are processed by the same simulator, and the GP hyperparameters β and ℓ are also selected heuristically on this setup. This creates a circularity risk: the magnitude of the gains may be specific to the calibration simulator. The paper should validate with cross-day or cross-environment testing, or at least report sensitivity of the headline gains to the calibration procedure.","section":"§IV.C and §V"},{"comment":"There is a mismatch between the stated objective and the algorithm. Equation (1) formulates a minimax objective (minimize the maximum assigned cost), and the accompanying text says the assignment minimizes the maximum cost. However, the Hungarian algorithm minimizes the sum of costs, not the bottleneck/makespan cost. Additionally, the constraints in (1) are malformed: the maximum is taken over c_{ij} x_{ij} without a sum over j, and the text then states both 'each task assigned to one robot' and 'each robot assigned to one task' in a way that is not cleanly encoded. The authors should either replace Hungarian with a bottleneck assignment solver for the stated minimax problem or re-state the objective as minimizing total cost and justify that choice.","section":"§III.A and §III.F"},{"comment":"The comparison does not isolate the contribution of the stochastic cost function. HATA differs from the path-based baseline in two ways: the A* planner avoids cells with human-presence probability above δ, and the bid cost includes the w1·η term in Eq. (3). Since the fitted weights are w0=1.15 and w1=0.95, the allocation cost is close to a weighted path length plus encounter probability, and the larger behavioral difference may be the δ-driven circumvention of high-density cells. No ablation is reported, e.g., MoD-avoiding A* with distance-only bids, or the HATA cost without δ avoidance. Without such ablations, the paper cannot attribute the observed gains to the proposed stochastic allocation cost rather than to path re-routing.","section":"§III.D–E and Table II"}],"minor_comments":[{"comment":"'MoDs Experimental results' should be 'MoDs. Experimental results'.","section":"Abstract"},{"comment":"The related-work paragraph contains a garbled sentence: 'Among decentralized approaches, auction-based methods, first This shift has led to interaction-aware planning methods...' This appears to be missing text and should be rewritten.","section":"§II"},{"comment":"In Eq. (3), the expectation of η_k is written as 'E[η_k] =: pi'; the symbol should be p_k (or a similarly declared probability), not the constant π, and it should be consistently subscripted.","section":"§III.E"},{"comment":"The notation in and around Eq. (1) is confusing: the objective and constraints are inconsistent, and the text repeats the assignment constraints in words rather than in a single clear mathematical form. This needs a careful rewrite even apart from the algorithmic mismatch noted above.","section":"§III.A"},{"comment":"Figures 2 and 4 lack error bars and have under-specified axes; for the main quantitative claims, the figure should include the same variability information requested for the tables.","section":"§V / Fig. 4"},{"comment":"The sentence 'a disk-shaped kernel of radius r = 10 cell around each tracked point' should state the physical radius (0.5 m if the grid resolution is 0.05 m) and explain how the kernel affects the probability computation in Eq. (2).","section":"§IV.A"}],"recommendation":"major_revision","confidential_remarks":"The manuscript header states 'Accepted final version. To appear in Proceedings of the IEEE.' If this paper is already accepted elsewhere, the editor should verify the submission status before considering it for this journal. My technical recommendation is based solely on the scientific content; the missing ablations, the minimax/Hungarian mismatch, and the absence of variability reporting are all addressable in a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: HATA is a reasonable incremental extension of human-aware MRTA, and the simulation effort is above average. But the 'up to 26%' improvements are not cleanly attributable to the stochastic cost, and the lack of variance reporting makes those numbers mush.\n\nWhat's new: combining Maps of Dynamics with auction-based allocation and a BO-tuned stochastic cost, evaluated in a coordination simulator with real pedestrian data from ATC. The held-out test day (Dec 12) is a real plus, and 1350 runs is more than most papers in this area bother with. If the gains hold up beyond this simulator, it's a useful step for intralogistics and delivery fleets.\n\nThe soft spots are real. The stress-test note is on target: HATA changes both the path planner (delta-based avoidance of high-density cells) and the bid cost (w1 * eta). There's no ablation separating those. Table II shows delta matters—at fleet size 15, delta=0.55 gives 281.1s, worse than HA-Alloc's 269.9s—so the 'up to 26%' is a favorable point in a swept space. The 90 runs per scenario should give you error bars or at least standard deviations; none are reported. The weights were fit on the same coordination simulator used for evaluation, so the effect size is simulator-specific. The paper's own limitations section notes the 15-robot cap due to coordination framework constraints, which is honest, but it also means the scalability claim is thin.\n\nThe 'first comprehensive evaluation' phrasing should be toned down. The prose has a few broken sentences (minor), but nothing that obscures the method.\n\nThis paper is for readers working on MRTA or human-aware navigation who want a concrete way to use long-term human motion models in allocation. It deserves a serious referee, but the revision needs an ablation (cost-only vs avoidance-only), variance reporting, and a sensitivity analysis on delta. Without that, the main claim is unproven.\n\nRecommendation: send it to peer review. It's an honest attempt with a plausible mechanism, but it needs the missing controls before the numbers can be trusted.","headline":"Solid incremental step in human-aware MRTA with an above-average simulation campaign, but the headline gains aren't isolated from path-avoidance effects and lack statistical support.","tokens_in":11186,"tokens_out":3097,"would_cite":true,"duration_ms":32787,"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":"Task allocation for robot teams improves when bids include the odds of encountering people along the route.","keywords":["multi-robot task allocation","Maps of Dynamics","human-aware planning","stochastic cost","pedestrian motion prediction","auction-based allocation","shared human-robot environments"],"falsifier":"Run HATA and a path-length auction on the same indoor floor plan with replayed or live pedestrian flows, re-tuning the weights on the actual robots; if the distance-only method's mission time is not higher by the predicted margin, or is lower, the central claim is not supported.","tokens_in":10326,"feed_emoji":"🤖","tokens_out":6036,"duration_ms":66784,"temperature":0.7,"pith_summary":"The paper is trying to establish that multi-robot task allocation in spaces where people move should use long-term pedestrian patterns, not just map geometry. It proposes HATA, which builds a time-indexed grid of pedestrian-presence probabilities from historical tracks and adds the expected encounter likelihood along each candidate path to the robot's bid. The two weights in that cost are tuned with Bayesian optimization inside a coordination simulator. In simulations that replay real pedestrian recordings, HATA reports mission times up to 26% shorter than a path-length auction baseline and up to 19% shorter than a prior human-aware method, with lower failure rates and waiting times. If this holds, fleet operators can cut delays by anticipating where people walk rather than reacting to them.","feed_headline":"Human-aware bidding cuts robot mission times by 26 percent","feed_subtitle":"Allocating tasks with maps of where pedestrians walk beats distance-only routing in shared spaces.","key_machinery":"The central object is the Map of Dynamics (MoD): a grid of cells, indexed by time of day, giving the probability that a person occupies a cell during a 30-minute interval. The cost identity is c_ij = Σ_k (w0·d_k + w1·η_k), where d_k is the Euclidean step length, η_k is a Bernoulli variable marking human presence, and w0, w1 are weights. A* planning uses MoDs with a threshold δ to route around high-probability cells; the resulting paths determine the cost bids; the Hungarian algorithm minimizes the maximum assigned cost; Bayesian optimization tunes the weights to match simulated mission times.","core_discovery":"HATA's central claim is that task allocation in human-shared spaces should treat route cost as a stochastic quantity. The cost of sending robot i to task j is the sum over path segments of a weighted Euclidean distance plus the expected presence of a pedestrian on that segment, derived from time-indexed Maps of Dynamics. The weights are found by Bayesian optimization in a coordination simulator, and the planner skips cells whose pedestrian-presence probability exceeds a threshold δ. In experiments replaying real shopping-mall pedestrian recordings with fleets of 5, 10, and 15 robots, the paper reports up to 26% lower mission completion time than a shortest-path auction baseline and up to 19%","pith_inferences":["This paper only evaluates centralized assignment; a natural extension the authors do not test is inserting the same MoD-based cost into decentralized auctions, which would show whether the gain comes from the cost model or from centralized assignment.","Because MoDs are indexed by time of day, updating them online from live pedestrian data is a testable way to handle drift; the paper uses a fixed historical map.","The simulator always gives pedestrians priority; with real negotiation or shared control, the optimal weights likely differ, and the advantage may shrink or grow depending on robot assertiveness."],"forward_implications":["Robots get tasks along routes that already avoid the densest pedestrian flows, so fewer local replans and stalls are needed.","The extra bidding cost from MoD lookup stays negligible compared with mission time, keeping the method scalable to larger fleets.","Operators can pick δ to favor speed (higher threshold) or safety (lower threshold) without changing the allocation machinery.","The reported gains appear across morning and afternoon time windows with different pedestrian densities, not at a single traffic level."],"supporting_citations":[{"why":"Source of Maps of Dynamics; HATA's stochastic cost relies on this spatio-temporal model of human movement patterns.","marker":"[7]"},{"why":"The HA-Alloc baseline that also integrates human motion into auctions; HATA is compared against it and claims lower computational overhead.","marker":"[6]"},{"why":"Supplies the real pedestrian tracking data used to build MoDs and to replay human trajectories during evaluation.","marker":"[24]"},{"why":"Coordination framework used to simulate robot-robot and robot-pedestrian interactions and measure mission, waiting, and failure metrics.","marker":"[27]"},{"why":"Bayesian optimization method used to tune the cost weights w0 and w1.","marker":"[19]"},{"why":"A* planner, extended with the δ threshold, generates the trajectories whose costs feed the assignment.","marker":"[18]"},{"why":"Hungarian algorithm solves the assignment from the robots' bids.","marker":"[23]"},{"why":"Auction-based path-length allocation used as the dynamics-agnostic baseline.","marker":"[28]"}],"fun_headline_variants":["Pedestrian-aware routing cuts robot mission time by 26%","Human-aware task allocation slashes multi-robot time 26%","Maps of pedestrian flow reduce robot task time up to 26%","Stochastic cost with human dynamics trims robot missions 26%","Robot task allocation with pedestrian maps speeds up 26%"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The two cost weights and the pedestrian-avoidance threshold are chosen using the same coordination simulator that is later used to measure the reported gains, and that simulator's delays are assumed to represent real human-robot encounters.","fun_headline_variants_meta":{"raw":{"variants":["Pedestrian-aware routing cuts robot mission time by 26%","Human-aware task allocation slashes multi-robot time 26%","Maps of pedestrian flow reduce robot task time up to 26%","Stochastic cost with human dynamics trims robot missions 26%","Robot task allocation with pedestrian maps speeds up 26%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000158,"raw_usage":{"total_tokens":1075,"prompt_tokens":767,"completion_tokens":308,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":219}},"tokens_in":511,"tokens_out":308,"duration_ms":4163,"temperature":1.0,"reasoning_tokens":219,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T15:29:45.278203+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run HATA and a path-length auction on the same indoor floor plan with replayed or live pedestrian flows, re-tuning the weights on the actual robots; if the distance-only method's mission time is not higher by the predicted margin, or is lower, the central claim is not supported.","supporting_citations":[{"cited_title":"Survey of maps of dynamics for mobile robots,","cited_arxiv_id":null,"evidence_quote":"Source of Maps of Dynamics; HATA's stochastic cost relies on this spatio-temporal model of human movement patterns."},{"cited_title":"Multiple robots avoid humans to get the jobs done: An approach to human-aware task allocation,","cited_arxiv_id":null,"evidence_quote":"The HA-Alloc baseline that also integrates human motion into auctions; HATA is compared against it and claims lower computational overhead."},{"cited_title":"Person tracking in large public spaces using 3-d range sensors,","cited_arxiv_id":null,"evidence_quote":"Supplies the real pedestrian tracking data used to build MoDs and to replay human trajectories during evaluation."},{"cited_title":"A loosely- coupled approach for multi-robot coordination, motion planning and control,","cited_arxiv_id":null,"evidence_quote":"Coordination framework used to simulate robot-robot and robot-pedestrian interactions and measure mission, waiting, and failure metrics."},{"cited_title":"Taking the human out of the loop: A review of bayesian optimiza- tion,","cited_arxiv_id":null,"evidence_quote":"Bayesian optimization method used to tune the cost weights w0 and w1."},{"cited_title":"A formal basis for the heuristic determination of minimum cost paths,","cited_arxiv_id":null,"evidence_quote":"A* planner, extended with the δ threshold, generates the trajectories whose costs feed the assignment."},{"cited_title":"The hungarian method for the assignment problem,","cited_arxiv_id":null,"evidence_quote":"Hungarian algorithm solves the assignment from the robots' bids."},{"cited_title":"Auction- based multi-robot routing","cited_arxiv_id":null,"evidence_quote":"Auction-based path-length allocation used as the dynamics-agnostic baseline."}],"review_version":1}