{"id":"a5cc2a88-908a-40aa-a533-35830092a9d6","arxiv_id":"2505.14580","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Tr-FMM is a Fast Marching Method variant that assigns a traversability score to each map region from obstacle density, movement, and goal alignment, and routes the robot through the best-scoring regions.","lead":"This robotics paper presents a path planner that divides an indoor map into regions and rates each region's safety based on moving obstacles, then steers the robot through less crowded regions while limiting detours. It is useful for robot navigation in public spaces like stores and airports, where going around a crowd can be safer and not much slower.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 5 assumes obstacle and robot speeds are similar, but the simulations use 0.2 m/s obstacles vs 0.5 m/s robot, so the reported safety gains and reduced deviations may be artifacts of a miscalibrated risk estimate.","rationale":"Both the safety benefit and the efficiency claim hinge on the risk classification in Eq. 5. If obstacle and robot speeds differ, the distance inequality no longer corresponds to who arrives first. The paper's own simulation parameters put obstacles at 0.2 m/s and the robot at up to 0.5 m/s, so the planner is conservative: it avoids regions that a faster robot could safely traverse, which lengthens paths. The results show Tr-FMM traveling longer distances than CBDB, which the authors attribute to a deliberate safety preference; the alternative explanation is that the distance heuristic overestimates risk. If a time-to-arrival comparison were used, the paths might be shorter with equal or better safety, undermining the claimed 'reducing unnecessary deviations.' The reader identified this same assumption, and the manuscript itself flags it, so this is not an artifact of our reading. A concrete test can settle it by varying obstacle speeds and comparing distance-based vs time-based classification.","tokens_in":19181,"tokens_out":9449,"duration_ms":85302,"concrete_test":"Re-run the dense and dispersed simulations from Section 4.1 under the same random seeds but with obstacle linear speeds of 0.2, 0.35, 0.5, and 1.0 m/s, and with a time-based variant of Eq. 5 that compares D_ri/v_robot to D+_doi/v_obstacle using measured obstacle speeds. If the time-based variant yields materially shorter paths or lower collision rates when obstacle speed differs from the robot's, the distance assumption in Eq. 5 is miscalibrated and the central safety benefit is not robust to the paper's own domain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3.1 defines traversability in Eq. 5: a region i is treated as non-risky when D_ri < D+_doi, a pure distance comparison. The authors state immediately after that 'the planner currently assumes obstacles move at a similar velocity to the robot, using distances.' This assumption is load-bearing for the central claim because the entire avoidance behavior is triggered by this inequality. In Section 4.1, the dynamic obstacles move at 0.2 m/s while the robot's maximum linear speed is 0.5 m/s, a 2.5x mismatch. For slower obstacles, Eq. 5 will classify as risky regions that the robot could actually reach first (because time = distance/speed, and the robot is faster), producing unnecessary detours and inflating traveled distance and mission time. The paper interprets these larger distances as a deliberate safety/efficiency trade-off, but they may simply be an artifact of over-penalizing regions that are not actually dangerous. Conversely, in real-world scenarios where pedestrians move faster than the robot, Eq. 5 will under-penalize risky regions, and the claimed collision avoidance could degrade. The manuscript acknowledges the assumption as a limitation but does not validate it, and the central claim of 'significant safety' with 'reducing unnecessary deviations' is precisely the quantity this miscalibration would distort.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Tr-FMM, a Fast Marching Method-based global path planner that discretizes the environment into regions, assigns each region a traversability value that combines goal deviation, obstacle occupation, and predicted obstacle dispersion, and then propagates the wavefront through regions of higher traversability while using a velocity map to avoid static and dynamic obstacles. The planner is evaluated in simulation against a basic FMM baseline, a social FMM (SFMM), and a crowd-based dynamic blockages method (CBDB), as well as in three real-world experiments with a Turtlebot, reporting fewer collisions and larger average distances to obstacles at modest additional path length and mission time. The authors claim that avoiding crowded regions directly during path search reduces unnecessary deviations and yields significant safety gains.","tokens_in":19463,"tokens_out":3889,"duration_ms":36612,"significance":"If the central claims are correct, the paper makes a useful practical contribution: a low-complexity, real-time-capable global planner that avoids congested areas without requiring detailed per-obstacle prediction. The real-world demonstrations, including a crowded hall with no collisions and sub-100 ms planning times, provide credible evidence of feasibility. The method is transparent and the underlying FMM machinery is standard, which supports reproducibility if the code becomes available. However, the quantitative safety and efficiency claims rest on a distance-based risk classification whose speed-matching assumption is not validated, and the simulation setup uses obstacle speeds that differ from the robot's speed by a factor of 2.5. Until this mismatch is addressed with either time-to-arrival reasoning or a sensitivity analysis, the reported improvements cannot be cleanly attributed to the method's design rather than to a miscalibrated risk model.","major_comments":[{"comment":"The risk classification in Eq. (5) compares the robot's distance to a region (D_ri) with the obstacles' estimated distance to that region (D+_doi), and the text immediately after the equation states that 'the planner currently assumes obstacles move at a similar velocity to the robot, using distances.' This assumption is load-bearing for the central claims of 'significant safety' and 'reducing unnecessary deviations.' In the simulations of Section 4.1, dynamic obstacles move at 0.2 m/s while the robot's maximum linear speed is 0.5 m/s, a 2.5x mismatch. For slower obstacles, Eq. (5) will classify as risky regions that the robot could actually reach first in time, producing detours that the paper then interprets as a deliberate safety/efficiency trade-off; for faster obstacles, the classification would under-penalize genuinely dangerous regions. Because the manuscript provides no sensitivity analysis over obstacle-to-robot speed ratios and no time-based validation, the reported gains in average distance and success rate may be at least partly artifacts of this miscalibration. Please either replace the distance comparison with a time-to-arrival estimate using tracked obstacle velocities, or add controlled experiments that vary the speed ratio and show that the qualitative results are preserved.","section":"3.3.1, Eq. (5)"},{"comment":"The region-seed suppression range is not specified. The paper states that 'all gradient values ∇D_so within the range of ∇Dso(xRi) are subtracted' when identifying region origins, but neither the formula for this range nor the value used in the experiments is reported. This range directly determines the number, size, and shape of regions, which in turn determines the traversability map and the resulting paths. Without this parameter, the method is not fully reproducible, and the sensitivity of the results to it is unknown. Please report the chosen range and include a brief sensitivity study, or, if the range is selected adaptively, provide the criterion used.","section":"3.2, Environment discretization"},{"comment":"The dispersion probability in Eq. (3) appears dimensionally inconsistent. Eq. (2) defines P_doi = Aτdoi / A_i for region i, but Eq. (3) then writes P^j_doi = (Aτdoi / A_j) · Pdoi, which reintroduces the same Aτdoi/A_i factor. This makes the dispersion probability depend on the destination region's own dynamism rather than solely on the source region j's movement, which is likely not the intended meaning. This directly affects the penalization D+_doi in Eq. (4) and hence the traversability in Eq. (5). Please clarify the intended definition and correct the notation.","section":"3.3.1, Eq. (3)"},{"comment":"The comparison with CBDB uses FMM instead of the A* algorithm originally employed in [28], with the claim that this 'guarantees that the robot maintains adequate distance from static obstacles.' However, no experiment is shown to verify that the FMM-based CBDB variant performs comparably to the original A*-based CBDB, so the comparison may implicitly penalize CBDB for a modification that is not validated. This is directly relevant to the paper's claim that Tr-FMM outperforms CBDB, and should be addressed with a calibration experiment or a discussion of the expected effect of this substitution.","section":"4.1.3, Comparison with other approaches"}],"minor_comments":[{"comment":"There is a typo in 'posisionts' that should read 'positions.'","section":"3.1, Online phase"},{"comment":"The notation in equations (1) and (5) uses ∇Dso(xτi) and ∇Dso(xi) without specifying whether these are normalized values or raw distances; please define the normalization more explicitly.","section":"2.2, FMM basics"},{"comment":"The real-world safety claim is based on a single 10-minute run with no collisions, and the authors note that people naturally avoided the robot. While this is a useful feasibility demonstration, it should be described as anecdotal rather than as a quantitative validation of safety.","section":"4.2, Experiment 3"},{"comment":"The statement 'The code will be available upon publication acceptance' does not provide a concrete artifact for reproducibility; please provide a repository link or a detailed parameter table in a revised version.","section":"1.2, Contributions"},{"comment":"The description of the wavefront insertion condition in Eq. (7) is terse; a short pseudocode description of the modified FMM update loop would improve clarity.","section":"3.3.2, Equations (6)-(7)"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant and timely problem, and the real-world experiments suggest the basic idea is viable. However, the central quantitative claim is tied to an explicit speed-matching assumption that is contradicted by the simulation parameters, and the parameter setting for region extraction is undeclared. These issues are fixable with additional experiments and clarifications, and I do not see an unfixable flaw in the conceptual design. I encourage the editor to view this as a conditional acceptance pending the sensitivity analysis and the clarification of Eq. (3)."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: Tr-FMM is a sensible, well-motivated extension of FMM for global planning in indoor dynamic environments. It discretizes the static map into regions from the gradient of distance to static obstacles, gives each region a traversability score that balances goal deviation, obstacle occupation, and a dispersion estimate, and changes the FMM wavefront insertion rule so region traversability is prioritized over raw distance. That combination is not in the prior FMM social planners or CBDB, and the paper's pipeline is clean: offline discretization, online traversability and wavefront propagation, and a complexity argument that keeps the whole thing at O(N_f log N_f). The experiments compare against SFMM and CBDB, include a replanning-once variant that speaks to the predictive value of the dispersion term, and the three real-world demos support feasibility. I found no evidence of cooking: the speed assumption is stated openly, and the replanning experiment is a fair attempt to isolate the dispersion effect.\n\nThe soft spots are concentrated in evaluation and one equation. First, the known speed assumption: Eq. 5 uses distances, not time-to-arrival, and the authors admit it assumes obstacle and robot speeds are similar. But in the simulations obstacles move at 0.2 m/s while the robot's max is 0.5 m/s. For slower obstacles, any region with D_ri >= D_doi is marked risky even when the robot would arrive first; that would systematically produce extra detours and larger obstacle distances, which are exactly the safety gains claimed. The limitation is acknowledged but not quantified, so the central quantitative claim is not yet supported as strongly as the prose suggests. Second, the dispersion probability in Eq. 3 looks muddled: it multiplies the dynamism term by an area ratio that appears to be from a different region, and the notation for P+_doi in Eq. 5 is never pinned down. That section needs a rewrite. Third, no code or data, and the bar charts have no error bars or significance tests; with 25 runs per condition, that's fixable but absent.\n\nThe paper is for someone working on FMM-based global planning or social navigation. The idea is worth a serious referee, and I'd engage with it. My recommendation: send to peer review, ask for a sensitivity analysis around the speed assumption, clearer dispersion equations, and the underlying data.","headline":"A plausible region-level traversability extension of FMM with a real speed-assumption problem in the evaluation; worth refereeing.","tokens_in":19978,"tokens_out":2767,"would_cite":true,"duration_ms":25878,"reading_group":"yes","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 robots in shared spaces navigate more safely by planning around whole crowded regions rather than around individual people, and that a traversability-weighted wavefront achieves this with modest detours.","keywords":["path planning","dynamic environments","traversability","Fast Marching Method","crowd avoidance","region discretization","mobile robot navigation","collision avoidance"],"falsifier":"Run the planner with full obstacle-position knowledge in one fixed indoor map, vary the obstacle speed from below to well above the robot's maximum speed while keeping the same seeded random motions, and compare collision rates and average distance to obstacles. If the distance-based rule of equation (5) is miscalibrated, obstacle speeds above the robot's speed should measurably increase collisions in regions the planner classified as non-risky ($D_{ri} < D^+_{doi}$), because those regions are judged safe only under the equal-speed assumption. A sharper test targets a single corridor region with $D_{ri} < D^+_{doi}$, with one obstacle accelerating toward it: a collision occurring before the robot's predicted arrival would falsify the risk classification directly.","tokens_in":18977,"feed_emoji":"🤖","tokens_out":8612,"duration_ms":69203,"temperature":0.7,"pith_summary":"This paper claims that in spaces shared with moving people, the safer global strategy is usually not to weave between individuals but to route the robot around entire crowded regions. Tr-FMM, the proposed planner, first discretizes the static map into regions with an adjacency graph, then assigns each region a traversability value combining three factors: deviation from the direct path to the goal, the disruptive occupation of dynamic obstacles within the region, and the probability that obstacles will disperse into neighboring regions. The Fast Marching wavefront is then propagated with region traversability as the primary priority, so the path prefers safer regions while the velocity map still avoids individual obstacles. Simulated and real-world experiments report larger average distances from obstacles and fewer collisions than two FMM-based baselines, at a modest cost in path length and mission time.","feed_headline":"Robot path planner dodges crowded regions, not just people","feed_subtitle":"A traversability-weighted wavefront keeps robots farther from moving obstacles with only modest detours.","key_machinery":"The load-bearing mechanism is a modified Fast Marching Method in which the wavefront's sorted set $W$ is ordered not by arrival-time gradient alone but by a two-key condition (equation 7): a new node outranks an existing one if it lies in a region of higher traversability, and only in ties does the lower gradient value decide. Traversability itself is assembled from three region-level quantities computed on the offline-built region graph $G_R$: deviation (shortest branch distances $D_{ri}$ and $D_{gi}$ from the robot's and goal's regions), occupation ($O_i$, the static-obstacle gradient $\\nabla D_{so}$ normalized along obstacle trajectories inside the region), and dispersion (dynamism $P_{doi}=A_{\\tau_{doi}}/A_i$ and the penalized obstacle distance $D^+_{doi}$). The offline discretization creates the regions by seeding from the maxima of $\\nabla D_{so}$ and letting colliding wavefronts from those seeds carve the boundaries, which makes each region encode the actual free-space shape around static obstacles.","core_discovery":"The central claim is that region-level traversability, rather than per-obstacle modeling, is the right unit for avoiding crowds in dynamic environments. The paper's object is the traversability value $T_{ri}$ of equation (5): for each region it combines the robot-goal distance through that region, $D_{rig}$, with an occupation-dispersion penalty $1-O_i P^+_{doi}$ that applies only when dynamic obstacles are expected to reach the region before the robot ($D_{ri} \\ge D^+_{doi}$). When the robot is expected to arrive first, the region keeps its full distance-based value and is treated as safe. From these region values, the wavefront propagation in equations (6)-(7) inserts new nodes with region traversability ranked ahead of gradient distance, which is what makes the planner choose a longer but safer route when a crowded region is avoidable, and fall back to less risky regions when it is not.","pith_inferences":["Because equation (5) compares distances rather than arrival times, the planner's risk classification is only calibrated when obstacles move at roughly the robot's speed; tracking obstacle velocities and switching to a time-of-arrival comparison is the natural next step, and the paper itself flags it as beyond its scope.","The paper's real-world Experiment 1 observes that a clutter-free room collapses into a single large region, which reduces routing choices; automatic subdivision of oversized regions into seeded sub-regions is a testable extension that would refine the granularity of crowd avoidance.","The same region-traversability machinery carries over to missions without a single goal, such as coverage, exploration, and multi-robot teams, by redefining the deviation term against the task's preferred sequence of regions rather than the direct path to one goal."],"forward_implications":["Robots using Tr-FMM will route around obstacle-dense regions whenever an acceptable detour exists, and through less-risky regions only when no free route remains, so crowd avoidance becomes a property of the global path rather than of reactive dodging.","Planning stays at the base FMM complexity: each replan costs $O(N_f \\log N_f)$, because region-traversability updates are constant time given the fixed offline region graph.","With only line-of-sight sensing, the planner retains most of its safety advantage, trading a modest increase in travel distance and mission time for larger average clearance from obstacles.","The dispersion term lets a single planned path remain effective without replanning, since the initial route already anticipates future occupation of adjacent regions.","In fully dispersed environments where no obstacle-free route exists, the planner degrades gracefully: it selects less-risky corridors and accepts higher collision exposure rather than failing outright."],"supporting_citations":[{"why":"Sethian's Fast Marching Method is the base wavefront propagation tool that Tr-FMM modifies.","marker":"[13]"},{"why":"Supplies the trick of using the static-obstacle gradient as the velocity map to keep safety distance, adopted directly by Tr-FMM.","marker":"[29]"},{"why":"SFMM, the FMM-based social path planner that serves as the primary baseline in the comparison experiments.","marker":"[26]"},{"why":"CBDB, the crowd-based dynamic blockage method that identifies crowded regions and is the second baseline; Tr-FMM extends it with dispersion and deviation terms.","marker":"[28]"},{"why":"Dijkstra's algorithm is used to build the region trees T_r, T_g, and T_R that supply all distances in the deviation and dispersion terms.","marker":"[20]"},{"why":"The Dynamic Window Approach is the local controller in every experiment, so its limited reactivity defines the collision metrics attributed to the planner.","marker":"[4]"},{"why":"Provides the FMM complexity bound used in the paper's O(N_f log N_f) complexity analysis.","marker":"[33]"}],"fun_headline_variants":["Path planner reroutes around crowds, not just obstacles","Region-based scores make robots avoid busy zones","Traversability-aware waves steer clear of crowds","Robot planner trades distance for crowd safety"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that dynamic obstacles move at about the same speed as the robot, so comparing distances rather than arrival times correctly decides which regions will be risky by the time the robot gets there.","fun_headline_variants_meta":{"raw":{"variants":["Path planner reroutes around crowds, not just obstacles","Region-based scores make robots avoid busy zones","Traversability-aware waves steer clear of crowds","Robot planner trades distance for crowd safety"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000136,"raw_usage":{"total_tokens":1105,"prompt_tokens":859,"completion_tokens":246,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":475,"completion_tokens_details":{"reasoning_tokens":189}},"tokens_in":475,"tokens_out":246,"duration_ms":3293,"temperature":1.0,"reasoning_tokens":189,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:31:02.537452+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the planner with full obstacle-position knowledge in one fixed indoor map, vary the obstacle speed from below to well above the robot's maximum speed while keeping the same seeded random motions, and compare collision rates and average distance to obstacles. If the distance-based rule of equation (5) is miscalibrated, obstacle speeds above the robot's speed should measurably increase collisions in regions the planner classified as non-risky ($D_{ri} < D^+_{doi}$), because those regions are judged safe only under the equal-speed assumption. A sharper test targets a single corridor region with $D_{ri} < D^+_{doi}$, with one obstacle accelerating toward it: a collision occurring before the robot's predicted arrival would falsify the risk classification directly.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Sethian's Fast Marching Method is the base wavefront propagation tool that Tr-FMM modifies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SFMM, the FMM-based social path planner that serves as the primary baseline in the comparison experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Dijkstra's algorithm is used to build the region trees T_r, T_g, and T_R that supply all distances in the deviation and dispersion terms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The Dynamic Window Approach is the local controller in every experiment, so its limited reactivity defines the collision metrics attributed to the planner."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the FMM complexity bound used in the paper's O(N_f log N_f) complexity analysis."}],"review_version":1}