{"id":"5e485451-98c2-444b-b693-8566e098b794","arxiv_id":"2606.19031","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A congestion-aware tour planner for service robots that samples human trajectories with CLiFF maps, models per-edge congestion in an MDP, and replans online to shorten tours in crowded environments.","lead":"This paper builds a robot tour planner that predicts where crowds will be using learned flow maps, then replans routes online to avoid congestion. It could help guide robots in museums, malls, and warehouses move faster through human-crowded spaces.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Speedup claim is confounded: Hamiltonian-path baseline lacks the wait and revisit actions available to the MDP, so the comparison does not isolate stochastic congestion modelling.","rationale":"The reader's weakest_assumption concerned the fidelity of CLiFF predictions and the synthetic simulation, but their rationale also noted that the 'sole baseline does not isolate the contribution of stochastic modelling'. I agree with that secondary point and believe it is actually the more load-bearing issue: even if the CLiFF predictions were ground truth, the current baseline's restriction to Hamiltonian paths (no revisits, no waiting) would likely still produce a speedup for LRTDP, because LRTDP is solving a strictly more flexible problem. This undermines the internal validity of the central empirical claim. The appropriate remedy is to strengthen the baseline rather than to abandon the framework, so the reader's CONDITIONAL verdict remains appropriate. I set agreement to 'partial' because the reader did not elevate the baseline confound to their weakest_assumption, even though it appears in the rationale.","tokens_in":11863,"tokens_out":7340,"duration_ms":77376,"concrete_test":"Run the same 40-simulation protocol with a new baseline that has exactly the MDP's action space: it can traverse any edge, revisit POIs, and wait, but it re-plans deterministically using current observed congestion as fixed edge costs (no CLiFF predictions or stochastic outcomes). For example, at each step solve a shortest-path/covering problem over the remaining POIs with the same allowed actions, using the currently observed congestion to set edge durations. If LRTDP still beats this enhanced baseline with p<0.05 across the four maps, the speedup is attributable to stochastic modelling; if not, the original comparison is confounded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim (Sec. V-B) is that LRTDP and LRTDP bounded beat the Hamiltonian-path solver. But the baseline is not solving the same problem. Problem 1 (Sec. IV) explicitly allows POIs to be revisited, and the tour MDP (Def. 7) includes a wait action and permits navigation to any outgoing edge, including edges to already-visited nodes. The baseline, by contrast, solves a Hamiltonian path over the remaining POIs using LKH 3 (Sec. V-A); this prohibits revisiting any POI and does not include waiting. Thus the LRTDP planner has a strictly larger action space: it can wait out congestion or revisit a node to avoid a crowded edge later, while the baseline cannot. The paper's own claim that this baseline 'allows us to evaluate the benefits of modelling stochastic crowd movement' is therefore not supported: the observed speedup may arise entirely from the extra actions rather than from reasoning over stochastic congestion. This is an internal mismatch between the problem definition and the baseline's representational power, not merely a difference of opinion with the broader literature.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an online tour planning framework for crowded environments. It uses CLiFF maps learned from human data to predict multiple trajectories for each observed human, maps predicted positions onto topological edges, computes time-dependent congestion probabilities over discrete congestion bands, and formulates tour planning as an SSP MDP with navigation and wait actions. The MDP is solved online with LRTDP and a bounded-time variant. Experiments on the ATC shopping mall dataset, replayed in a synthetic simulator across four topological map sizes and three congestion-band settings, report statistically significant reductions in tour duration compared to a Hamiltonian-path baseline.","tokens_in":12131,"tokens_out":7346,"duration_ms":71679,"significance":"The formal construction is coherent: the MDP formulation is standard, the congestion band abstraction is a reasonable way to bound branching, and the use of CLiFF-LHMP predictions separates the learned motion model from the planning objective (no circularity). The paper also ships an open-source implementation (link removed for anonymity) and includes statistical tests over 40 runs per condition. If the empirical claim were properly supported, the framework would be a useful contribution to service-robot tour planning. The main weaknesses are that the evaluation is entirely synthetic, the cost model is hand-designed, and the baseline is representationally weaker than the MDP, so the specific claim that stochastic congestion modelling drives the speedup is not yet established.","major_comments":[{"comment":"The central empirical claim that LRTDP beats the Hamiltonian path solver is confounded by an asymmetric action space. Problem 1 explicitly allows POIs to be revisited, and Def. 7 includes a wait action. The baseline uses LKH 3 to synthesize a Hamiltonian path over the remaining POIs, which prohibits revisits and has no wait action. The MDP can wait out congestion or defer an edge by visiting other nodes, while the baseline cannot. Thus the sentence in Sec. V-A that this baseline 'allows us to evaluate the benefits of modelling stochastic crowd movement' is unsupported. Please add a baseline with the same action space (e.g., re-planning with wait and revisit allowed but using static congestion estimates), or ablate the tour MDP without wait/revisit, before claiming the speedup is due to stochastic congestion reasoning.","section":"Sec. V-A and V-B; Problem 1; Def. 7"},{"comment":"The congestion probabilities used in Eqs. (3)-(4) are computed from m=10 sampled CLiFF-LHMP trajectories per human. The CLiFF map is trained on half of the ATC dataset and evaluated on the other half, but no prediction-accuracy measure is reported for the trajectory or edge-occupancy predictions. If the predictions are inaccurate, the MDP's transition probabilities are mis-specified and the simulated speedup may not transfer to real environments. Please report a prediction-error metric on the held-out data (e.g., displacement error or edge-occupancy error) and, ideally, a sensitivity analysis with respect to m.","section":"Sec. IV-B, Eq. (1), Sec. V-A"},{"comment":"The simulator's traversal duration is a hand-designed formula: 'edge distance multiplied by the robot speed' plus a fixed 10-second penalty per human encountered. This cost model is not derived from the ATC data or from any empirical relationship between crowd density and robot traversal time, and the paper does not specify how the duration distributions ρ(e,c_j) are set in the experiments. The quantitative gains (e.g., 109s vs 132s on the 26-node map) are therefore sensitive to an arbitrary constant. Please specify ρ explicitly and include a sensitivity analysis over the penalty magnitude to show that the qualitative conclusions are robust.","section":"Sec. V-A; Def. 1; Eq. (4)"}],"minor_comments":[{"comment":"Typo: 'the edge distance multiplied by the robot speed' should be 'divided by the robot speed'. Also 'ten 50second trajectories' should be 'ten 50-second trajectories'.","section":"Sec. V-A"},{"comment":"The location l_i is written as l_i ∈ R, but a 2D location should be l_i ∈ R². Please correct.","section":"Def. 4"},{"comment":"The expression Ψx_i(t) ∈ R_e is a boolean test; use indicator notation (e.g., 1[Ψx_i(t) ∈ R_e]) to avoid ambiguity. Similarly, in Eq. (3), '1e=wait' is unclear; write 1_{e=wait}.","section":"Eq. (1)"},{"comment":"The p-values are not corrected for multiple comparisons. Although most are very small, a multiple-comparison correction or a footnote explaining the family-wise error rate would strengthen the statistical claims. Also, 'Values in bold' is not visible in the plain-text table; mark the significant entries explicitly.","section":"Table III"},{"comment":"The assumption of external sensors that observe all humans is stated, but the impact of detection noise, occlusion, or missed detections on the congestion model is not discussed. Please add a sentence on this limitation or a reference to work that handles perception uncertainty in this setting.","section":"Sec. IV-A"},{"comment":"The conclusion says 'demonstrate efficacy on a real-world crowd dataset', but the evaluation is a synthetic simulation that only uses the ATC data to train the CLiFF map and replay human movements. A limitations paragraph acknowledging the synthetic cost model and the absence of physical robot experiments would be appropriate.","section":"Sec. VI"}],"recommendation":"major_revision","confidential_remarks":"The paper is well written and the formal MDP part is sound, but the baseline confound in Sec. V-B is load-bearing for the central claim. If the authors cannot add a fair baseline (or ablate the wait/revisit actions), the claim about stochastic congestion should be softened. The arXiv header already says the paper is accepted at IROS 2026, which is unusual for a submission, but I do not treat it as a technical issue. I would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know about this paper: it's a clean integration of CLiFF-based human trajectory prediction into an online SSP-MDP for tour planning, with a plausible synthetic evaluation on the ATC mall dataset. The strongest advertised result—LRTDP beating the Hamiltonian path baseline—does not actually hold up as a demonstration of stochastic congestion modeling, because the baseline is solving a different problem.\n\nWhat's genuinely new: the authors take their own multi-robot congestion MDP machinery [35], swap in CLiFF-LHMP human predictions, and show the resulting planner can re-plan online. The MDP definition is coherent, the congestion-band idea is practical, and the use of a train/test split of real human data is a step up from pure simulation. The planning-time results—median about a second on the largest map—suggest the approach is scalable enough for online use. The statistical tests are a nice touch.\n\nWhere it goes soft: the baseline (Sec. V-A) solves a Hamiltonian path over remaining POIs using LKH 3. It cannot wait, and it cannot revisit a POI. But the MDP (Def. 7) has a wait action and allows navigation to any outgoing edge, including edges to already-visited nodes. Problem 1 explicitly allows revisits. So the comparison does not isolate the benefit of stochastic reasoning; the extra action space alone could explain the speedup. The paper's claim that this baseline 'allows us to evaluate the benefits of modelling stochastic crowd movement' is therefore not supported. That's a real confound in the headline result.\n\nBeyond that, the evaluation is entirely synthetic: perfect sensing of all humans, a hand-set 10-second penalty per human on an edge, and no reported accuracy measure for the CLiFF-LHMP predictions themselves. The code/data link is redacted, which makes reproduction harder. None of these would be fatal on their own, but together they leave the practical impact unproven.\n\nWho should read it: people working on task-level planning for service robots in crowded spaces. It's a within-subfield contribution, but a useful one to know about. I'd send it to a serious referee, but with a request to fix the baseline—either give the baseline the same wait/revisit options, or run an ablation that isolates stochastic modeling from action-set flexibility. As is, the central claim is conditional.","headline":"Nice integration of CLiFF-human prediction into an online SSP-MDP tour planner, but the headline speedup over the Hamiltonian baseline is confounded because the baseline lacks the wait and revisit actions the MDP has.","tokens_in":12637,"tokens_out":3563,"would_cite":false,"duration_ms":33230,"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 an online tour planner that explicitly reasons over stochastic human congestion completes robot tours faster than a baseline that assumes congestion is static, with statistically significant speedups across map sizes","keywords":["congestion-aware tour planning","CLiFF map","Markov decision process","LRTDP","human motion prediction","receding horizon planning","crowd navigation","service robots"],"falsifier":"Run the planner on the held-out half of the ATC dataset with a perfect-forecast oracle as an upper bound and a static-congestion baseline as a lower bound; if the CLiFF-based planner's advantage over the static baseline disappears when predicted trajectories are replaced with ground-truth future positions, the claimed benefit is an artifact of the simulator's cost model. Alternatively, deploy the planner on a real robot in the mall and compare measured tour durations against a non-congestion-aware route under matched conditions.","tokens_in":11744,"feed_emoji":"🤖","tokens_out":3237,"duration_ms":32050,"temperature":0.7,"pith_summary":"The paper tries to establish that a mobile service robot can plan faster tours through crowded spaces by treating congestion as a stochastic, time-varying quantity rather than a static snapshot. It builds this reasoning on CLiFF maps, which predict human trajectories from observed motion, and embeds the resulting congestion probabilities into a Markov decision process that is solved online with a real-time heuristic search. In simulations replaying real shopping-mall crowd data, both the full and time-bounded versions of the planner complete tours faster than a Hamiltonian path solver that uses a fixed congestion estimate, and the differences pass a statistical significance test. The practical message is that tour planning should be coupled to human-motion prediction, and that even a coarse congested/uncongested split may capture most of the benefit.","feed_headline":"Crowd-aware replanning shortens robot tours in mall simulations","feed_subtitle":"A planner that re-predicts human flow at each step outruns a fixed-route baseline in real-world crowd data.","key_machinery":"The central mechanism is a pipeline that converts raw human observations into time-dependent congestion probabilities over topological edges: CLiFF-LHMP samples multiple trajectories per human, each trajectory is mapped onto the rectangles around edges, a Poisson-binomial distribution gives the probability of q humans on an edge at time t, and these counts are aggregated into congestion bands. The bands feed a stochastic-shortest-path MDP (the tour MDP), whose states track the robot's node, current time, and visited POIs; costs are expected traversal durations under each band. The MDP is solved online by LRTDP with an MST-based admissible heuristic, and the robot executes one action before r","core_discovery":"The central discovery is that stochastic congestion modeling pays off at the tour-planning level: an MDP whose edge costs are time-dependent distributions over the number of humans on each edge, updated at every step from CLiFF-map trajectory predictions, produces tours that are faster to execute than those of a baseline that solves a Hamiltonian path problem under a one-time congestion estimate. The paper demonstrates this on four topological maps built from ATC shopping-mall data, with two, five, and eight congestion bands, and shows that the online time-bounded planner performs nearly as well as the converged planner, while median first-step planning time stays around one second even on t","pith_inferences":["The core formulation is domain-generic: the same MDP construction applies to any human-populated environment, from malls and museums to fulfilment centres, provided a map of dynamics and a topological graph are available.","If the binary-band result holds beyond these simulations, simpler congestion-aware planners (even hand-tuned rules based on predicted crowd levels) might capture most of the benefit, potentially reducing computation further.","The fidelity of the CLiFF-LHMP trajectory predictions is the load-bearing external input; a direct measure of prediction accuracy against the held-out half of the ATC data would clarify how much of the speedup derives from prediction quality rather than from the planning formulation itself.","The 10-second per-human penalty and the assumption that every human is observed are simulation simplifications; a real deployment would need to quantify how trajectory prediction errors and partial sensor coverage degrade the expected gains."],"forward_implications":["Tour plans that explicitly reason over congestion produce shorter expected tours than static-congestion Hamiltonian paths in the tested shopping-mall scenarios, statistically significant at p=0.05.","A time-bounded version of LRTDP (3 seconds per planning step) achieves nearly identical tour durations to a converged LRTDP, suggesting the approach is usable online.","Even with coarse congestion bands (congested versus not), execution times are similar to finer-grained band settings, implying a binary congestion model may be sufficient in practice.","Planning time scales reasonably: median first-step planning time is about one second on the largest map tested (26 nodes), suitable for online replanning.","The framework is reactive: replanning after each action lets the robot adapt to humans entering or leaving the environment during execution."],"fun_headline_variants":["Crowd-aware replanning cuts robot tour times in malls","Probabilistic tour planner flexes with crowd flow, beats static route","Robot tour planner uses live crowd estimates to shorten routes","Online MDP replanning trims robot tours in crowded environments","Congestion-adaptive robot tours outpace fixed plans on mall data"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The planning advantage rests on the assumption that CLiFF-LHMP trajectories, sampled from a map learned on one half of the shopping-mall data, faithfully predict where humans will be on the held-out half; the paper reports no accuracy measure for these predictions, and the simulator further assumes every human is observed and each encounter costs exactly ten seconds.","fun_headline_variants_meta":{"raw":{"variants":["Crowd-aware replanning cuts robot tour times in malls","Probabilistic tour planner flexes with crowd flow, beats static route","Robot tour planner uses live crowd estimates to shorten routes","Online MDP replanning trims robot tours in crowded environments","Congestion-adaptive robot tours outpace fixed plans on mall data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000209,"raw_usage":{"total_tokens":1212,"prompt_tokens":679,"completion_tokens":533,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":423,"completion_tokens_details":{"reasoning_tokens":443}},"tokens_in":423,"tokens_out":533,"duration_ms":5925,"temperature":1.0,"reasoning_tokens":443,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T10:54:17.566008+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the planner on the held-out half of the ATC dataset with a perfect-forecast oracle as an upper bound and a static-congestion baseline as a lower bound; if the CLiFF-based planner's advantage over the static baseline disappears when predicted trajectories are replaced with ground-truth future positions, the claimed benefit is an artifact of the simulator's cost model. Alternatively, deploy the planner on a real robot in the mall and compare measured tour durations against a non-congestion-aware route under matched conditions.","supporting_citations":[],"review_version":2}