{"id":"50907a6c-4738-423b-888a-5546c4e10594","arxiv_id":"2502.04837","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A social navigation framework clusters people into groups, models group proxemics as a superposed magnetic-dipole vector field, computes optimal observation positions, and plans cruise paths with vector-field-guided RRT*.","lead":"The paper builds a robot navigation system that recognizes human groups and steers the robot to good viewing spots, using a magnetic-dipole vector field to model social space. It then plans cruising paths among groups with a modified RRT* planner, and tests the pipeline in simulation and on a real service robot.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The OOP computation in Eq. (13)-(16) is an implicit, possibly unsolved equation with ambiguous sign (β=-32) and uncalibrated constants (α=-615); if Eq. (13) has no root or multiple roots for the reported configurations, the optimal-observation-position claim and the VMD-RRT* cost are unsupported.","rationale":"The reader's CONDITIONAL verdict is appropriate. The framework is coherent, and the VMD-RRT* versus RRT/RRT* comparison in Table III is genuine independent evidence that the planner wrapper has some effect. However, the OOP step is where the social model enters the pipeline, and it is underdetermined as written. My concern is not that the magnetic-dipole model disagrees with existing HRI consensus; it is more concrete: Eq. (13) is an implicit nonlinear equation with no shown existence, uniqueness, or numerical solution, and the sign of β together with the choice of α and β has direct geometric consequences for where the robot stops. A residual and sensitivity check is feasible and would settle whether the reported OOPs actually satisfy Eq. (13) or are artifacts of an unspecified solver. If the check fails, the paper should provide a well-posed OOP definition with calibration or soften the optimality claim; if it passes, the conditional acceptance can proceed. I therefore keep the reader's verdict unchanged while sharpening the condition that should be imposed.","tokens_in":19275,"tokens_out":7208,"duration_ms":78467,"concrete_test":"First fix the intended sign convention for Eq. (14), since the printed β=-32 makes the amplitude negative. Then, for every group configuration in Fig. 8 and the static experiment of Fig. 12, solve Eq. (13) by root-finding or fixed-point iteration over the 14 m by 14 m map and report the residual ||V_gk(O_k)+ξ_k(O_k)||, the number of roots found, and the solver used. Finally, re-run the full pipeline with α=-615 replaced by -307.5 and -1230, and separately with β=-32 replaced by -16 and -64, holding all other parameters fixed; if the OOPs move by more than about 1 m, or if residuals are not below 1% of |β|e^{-n}/n, the social model constants are load-bearing and the 'optimal' claim is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on Eq. (13)-(16) actually defining a unique, meaningful OOP, and on constants α=-615 and β=-32 being tied to human proxemics. Neither is secured in the manuscript. Eq. (13) is implicit in O_k twice: the field V_gk depends on O_k directly, and ξ_k's argument is the average bearing from O_k to the group members (Eq. (15)-(16)). The paper gives no existence or uniqueness argument, no solver description, and no residuals for the OOPs shown in Fig. 8 or Fig. 14. The problem is not merely formal: with Eq. (14) as written, β=-32 makes Amp(ξ_k) negative, so the vector equation's sign is ambiguous; for a single person at heading 0, solving Eq. (13) under the natural sign convention places the root behind the person at about 4.7 m, an implausible observation position, while the opposite sign convention places it in front at the same distance. For multi-person groups the dipole field's angular structure can create zero or multiple roots. Since no calibration against Hall's proxemics or human perceptual data is reported, and no sensitivity analysis over α and β is provided, the OOP grid and the social cost in VMD-RRT* are not shown to be stable or meaningful. The comparison to RRT/RRT* is independent evidence, but it validates the planner wrapper, not the proxemics model.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an online robot motion planning framework that blends social proxemics of human groups into navigation. It introduces a graph-based clustering method combining social relevance and spatial confidence (Section II-A), a magnetic-dipole vector-field model of individual and group proxemics (Section II-B), a method to compute optimal observation positions (OOPs) via a vector balance equation (Eq. (13)), and a sampling-based planner called VMD-RRT* that uses the proxemics field in its cost function (Section II-C). The authors validate the approach with simulations in a 14 m x 14 m map with static and moving pedestrians, and with real-robot experiments using a Vicon motion capture system and the Xiaopang robot in static and dynamic social scenarios. They report high group recognition, successful OOP arrival, and improved time/path/node statistics compared with RRT and RRT*.","tokens_in":19593,"tokens_out":2762,"duration_ms":31897,"significance":"If the proxemics model and the OOP computation are sound, the paper offers a useful way to make robot navigation interaction-aware rather than obstacle-avoidance-only, and the inclusion of real-robot experiments with dynamic group reconfiguration is a strength. The paper also provides a reproducible simulation setup and a direct planner comparison (Table III) that gives independent evidence for the value of the hierarchical planning wrapper. However, the paper's central claim depends on the magnetic-dipole proxemics model and the OOP equation being meaningful and well-posed, and that is not currently established: the model constants are asserted rather than calibrated, the implicit OOP equation is not shown to have a unique solution, and the main success metric (Eq. (25)) is partly self-referential. These gaps make the contribution conditional rather than fully convincing.","major_comments":[{"comment":"The OOP is defined as a solution of the implicit vector equation V_gk(O_k) + xi_k = 0, but the amplitude of xi_k in Eq. (14) is negative for the reported beta = -32, making the sign convention ambiguous. Moreover, Arg(xi_k) in Eq. (15) depends on O_k through Eq. (16), so Eq. (13) is a nonlinear implicit equation in O_k. The paper provides no existence or uniqueness argument, no description of the numerical solver, and no residuals for the OOPs displayed in Figs. 8 and 14. As written, the 'optimal observation position' is not actually shown to exist, be unique, or be computable for the reported configurations. Please provide a solver, a sign convention (e.g., define Amp(xi_k) with absolute value or change beta), and a numerical check of existence/uniqueness over the tested group configurations.","section":"II-B, Eq. (13)-(16)"},{"comment":"The constants alpha = -615, beta = -32, a = 5.102, b = 0.748, c = 0.087, r = 0.05, and the cost weights in Eq. (20) are asserted without calibration against human proxemics data or any perceptual ground truth. Since the social cost in VMD-RRT* is computed from V_gC, and the OOP is defined as a zero of a field with these constants, the reported paths and OOPs are not yet shown to reflect human social preferences. At minimum, provide a sensitivity analysis over alpha and beta, and ideally compare the resulting OOP distances and approach orientations with established proxemics ranges (e.g., Hall's proxemics or the social interaction field model of Ref. [12]).","section":"II-B, Table I"},{"comment":"The Arrive Rate metric A_r measures error against the OOP and the damping orientation Arg(xi_k) that the model itself generates, so high A_r largely demonstrates self-consistency of the optimization, not external validity of the proxemics model. To support the claim that the robot reaches socially appropriate observation positions, please add an independent evaluation, for example human-ratings of the robot's approach, comparison with a random or heuristic OOP baseline, or a measure based on the group gaze center gamma_k defined in Eq. (19).","section":"IV-B, Eq. (25)"},{"comment":"The comparison with RRT and RRT* in Table III shows that VMD-RRT* reduces path length and nodes, but the practical framework also uses a TSP-based visiting sequence, informed sampling in an ellipse, and hierarchical sub-goal replanning (Section II-C3). These factors, rather than the proxemics field itself, may drive the improvement. Please include an ablation that separates the effect of the vector-field cost from the effect of the hierarchical/TSP wrapper, so the reader can see how much the social proxemics field contributes.","section":"III-B, Table III"}],"minor_comments":[{"comment":"The abstract claims 'promising performance on group recognition accuracy', but no quantitative clustering accuracy metric is reported; the dynamic scenario in Fig. 14 is shown only qualitatively. Please add a clustering evaluation (e.g., precision/recall against labeled groupings).","section":"Abstract and Section III"},{"comment":"Equation (1) is difficult to parse: the placement of the norm in the denominator and the layered product are unclear. Please rewrite with explicit parentheses and define all terms before use.","section":"II-A, Eq. (1)"},{"comment":"The function psi(cos theta) and the angles theta_A and theta_B are defined only in prose. Please give precise definitions of the angle arguments in the equations, and note that psi is not differentiable at 0 if that matters for the clustering.","section":"II-A, Eq. (2)-(3)"},{"comment":"In Eq. (24), 'ln V_gC(sigma)' appears to take the logarithm of a vector, while the text says 'ln is logarithmic function with base e'. If the intended quantity is the log-magnitude, please write it explicitly as ln ||V_gC(...)||_2.","section":"II-C, Eq. (24)"},{"comment":"Algorithm 1 has duplicate line numbers 17, and the 'CollisionFree(s_new, s_min)' check only tests two states rather than the edge between them. Also, the variable 'near' in Line 9 is introduced without definition; it should be s_nearest or s_near.","section":"Algorithm 1"},{"comment":"Equation (25) contains a likely typo: the second term is written '(x_y - O_y)^2' instead of '(y_r - O_y)^2'. Please correct.","section":"IV-B, Eq. (25)"},{"comment":"The column header says 'Success Rate' but the entries are fractions such as '10/10'. This is understandable, but please state explicitly that the denominator is the number of repeated trials; currently that is only implied in the text.","section":"Table II"},{"comment":"The gaze-point constant r = 0.05 in Table I seems very small compared with typical interpersonal distances; please clarify its units and justify the value, or state that it is a normalized quantity.","section":"II-B, Eq. (17)-(19)"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant problem in social robot navigation, and the real-robot experiments are a positive feature. However, the central proxemics model and the OOP computation are not yet established as meaningful or well-posed, and the main success metric is partly self-consistent by construction. I recommend major revision rather than rejection because the issues are addressable with additional numerical analysis, calibration, and ablation experiments. One additional editorial concern: the manuscript contains several typos and inconsistent notations that should be corrected before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: I agree with the conditional verdict. The package is genuinely new: they use a magnetic-dipole field for individual proxemics, superpose it into group and global fields, define an 'optimal observation position' from the damped field, and feed that vector field into an RRT* variant. The group clustering is close to Zhou et al.'s social interaction field, so the novelty is in the integration rather than the individual pieces. The real-robot dynamic experiments are the best evidence. The comparison to RRT/RRT* in Table III is independent of the proxemics model and favors VMD-RRT* on path length, node count, and planning-time stability. That is a real result, though from one setup. The citation pattern is reasonable: Hall, Zhou's interaction field, and RRT* baselines are all there.\n\nThe soft spots are all around the proxemics model and the OOP. Eq. (13) is implicit in the OOP and has no existence, uniqueness, or solver discussion. With beta = -32, Amp(xi) is negative, so the sign convention matters; the stress-test worked example (a single person at heading 0) puts the root behind the person under the natural convention, which is not a sensible observation position. Multi-person fields can plausibly give zero or multiple roots, and nothing in the paper rules that out. The constants alpha = -615 and beta = -32 are asserted, not calibrated against Hall's proxemics or any human perceptual data, and there is no sensitivity analysis. That matters because the same field drives the VMD-RRT* social cost.\n\nThe 'optimal' label is not earned. There is no objective function; Eq. (13) just defines a zero of a damped field. The Arrive Rate in Eq. (25) measures error relative to the OOP and damping orientation produced by the same model, so a high value mostly shows self-consistency, not social validity. The abstract's claim of 'group recognition accuracy' has no reported accuracy metric behind it, and the static experiment has no baseline. The independent RRT/RRT* comparison supports the planner contribution, not the proxemics model.\n\nI would send this to a serious referee. The framework is clear, the experiments are real, and the planner comparison is useful. Reviewers should require an existence/uniqueness or numerical-solver analysis for Eq. (13), calibration or sensitivity analysis for alpha and beta, and a reported clustering accuracy metric. If those survive, this becomes a solid contribution to social navigation; right now it is a credible framework with overclaimed components. For readers, it is worth knowing about, but not yet a model to build on.","headline":"A real-robot social navigation pipeline with a genuinely new proxemics representation, but the OOP computation and model constants need validation before the 'optimal' and accuracy claims hold up.","tokens_in":20185,"tokens_out":6385,"would_cite":true,"duration_ms":68602,"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":"A robot can cruise among human groups by navigating a vector field assembled from each person's social space.","keywords":["social robotics","proxemics","group clustering","magnetic dipole model","optimal observation position","sampling-based path planning","human-robot interaction","social navigation"],"falsifier":"A direct test would be a perception experiment: for groups of two to four people at varied spacings and headings, compare the predicted optimal observation position from Eq. (13) with the positions human observers actually choose for a robot to approach, or with questionnaire ratings of approach comfort; if the predicted point consistently falls outside the preferred interaction region, or if small heading changes (around 10 degrees) move the predicted OOP by more than about a meter, the dipole superposition is not capturing group proxemics.","tokens_in":18977,"feed_emoji":"🤖","tokens_out":13393,"duration_ms":115381,"temperature":0.7,"pith_summary":"This paper claims that a service robot can navigate among human groups in a socially aware way by representing each person's personal space as a two-dimensional vector field shaped like a magnetic dipole and adding these fields together for groups and the whole scene. It introduces a graph-based clustering step that groups people using both social relevance and a spatial-confidence score, then defines the best place to stand near each group, the optimal observation position (OOP), as the point where the group's total field is exactly cancelled by a social damping vector. The planner, called VMD-RRT*, steers the robot with a cost that rewards moving along the field while keeping out of the groups' social gaze space. In simulations and on a physical service robot with volunteers, the robot re-clusters people as the scene changes and visits each group's OOP at a socially comfortable distance. The paper's point is that proxemics can be a planning input for interaction, not just an avoidance constraint.","feed_headline":"Proxemics field steers a robot to the right spot to meet social groups","feed_subtitle":"A dipole-field model of personal space tells a robot where to stand and how to keep a polite distance.","key_machinery":"The load-bearing object is the magnetic-dipole social proxemics field: a two-dimensional vector field $\\mathcal{M}$ defined in Eqs. (7)-(8) with coefficient $\\alpha=-615$, rotated by the person's heading to form the individual's social space $\\boldsymbol{S}$. Group and scenario maps are built by linear superposition ($\\boldsymbol{V}_{g_k} = \\sum_j \\boldsymbol{S}_j$, global field $\\boldsymbol{V}_{gC} = \\sum_k \\boldsymbol{V}_{g_k}$). The optimal observation position is defined by the zero-damping condition $\\boldsymbol{V}_{g_k}(O_k) + \\boldsymbol{\\xi}_k = 0$, where $\\boldsymbol{\\xi}_k$ has amplitude $\\beta e^{-n}/n$ and direction given by the mean angle from the group plus $\\pi$. This single field structure does double duty: it produces the OOP grid used as goals, and it supplies the direction and magnitude costs $\\mathcal{F}_{dir}$ and $\\mathcal{F}_{mag}$ inside VMD-RRT*, so the same representation both selects where to stand and shapes the path to get there.","core_discovery":"The central discovery is that group proxemics can be reduced to a single continuous two-dimensional vector field whose structure tells a robot both where to stand and how to get there. Each person contributes a magnetic-dipole field $\\mathcal{M}$ (Eqs. (7)-(8)) rotated by their heading, the group field is the linear superposition $\\boldsymbol{V}_{g_k} = \\sum_j \\boldsymbol{S}_j$, and the robot's target, the OOP, is the point where $\\boldsymbol{V}_{g_k}(O_k) + \\boldsymbol{\\xi}_k = 0$ with damping amplitude $\\beta e^{-n}/n$ set by group size. The same field, summed over all groups, is fed into an RRT*-style planner whose edge cost penalizes path length, field-direction mismatch, and field magnitude, so the robot cruises between OOPs without entering social gaze spaces. The authors report that this framework, tested in simulation and with a real service robot, recognizes groups, re-plans when groups merge or move, and keeps a distance above the 0.4 m threshold while visiting all groups.","pith_inferences":["Going beyond the paper, the same zero-damping condition could be read as a general social-equilibrium viewpoint generator, applicable to multi-party human-robot interaction or camera viewpoint selection for group photography, not only robot cruising.","A testable extension the paper does not perform is calibration: fitting $\\alpha$, $\\beta$, and the Weibull parameters $a,b,c$ to measured human comfort ratings or perceptual grouping judgments would turn the model from an analogy into a predictive proxemics theory; the current constants are fixed without such data.","Because the framework assumes full pose information from a motion capture system, a practical next step is to replace that input with onboard person detection and tracking; the sensitivity of OOP positions to heading noise would then be the key question.","The TSP-plus-sampling decomposition suggests a natural online variation: predict group motion and re-solve the visiting sequence over a short horizon, which would reduce redundant observations like the duplicate visit to $\\mathcal{G}_2'$ seen in one dynamic experiment."],"forward_implications":["A robot using this framework can treat people as interaction targets rather than obstacles, so the same proxemics field that keeps it at a respectful distance also tells it where to stop and initiate contact.","Because the clustering and field are recomputed online, a change in group composition (people joining, leaving, or moving) propagates directly into a new OOP set and a new cruise path without switching to a separate avoidance mode.","The reported comparison against RRT and RRT* on the physical platform indicates that field-guided sampling can reach the same observation goals with shorter path length and fewer explored nodes (about 24 m and 6,500 nodes for VMD-RRT* versus 26 m and 20,000 nodes for RRT*).","The design space has a documented trade-off: enlarging the robot motion space radius speeds planning but lowers success in dynamic scenes (from 10/10 at $r_{rob}=1.9$ m to 4/10 at $r_{rob}=2.7$ m with two moving individuals), so the radius must be tuned to the expected scene dynamics."],"supporting_citations":[{"why":"Supplies the social interaction field model that motivates the social-relevance term in group clustering and supports identifying static and dynamic groupings.","marker":"[12]"},{"why":"Offers the existing asymmetric-Gaussian representation of individual and group zones that the proposed vector-field model is positioned against.","marker":"[21]"},{"why":"Contributes the idea of continuous pedestrian space functions, the conceptual basis for encoding proxemics as a field.","marker":"[20]"},{"why":"Provides artificial potential field theory, which motivates building the guidance field by superposition of component fields.","marker":"[31]"},{"why":"Supplies the planetary magnetic-field analogy that the dipole form of the individual proxemics model is drawn from.","marker":"[35]"},{"why":"Provides the classical RRT* algorithm and the Steer function that VMD-RRT* inherits and extends with the social vector field cost.","marker":"[26]"},{"why":"Supplies the TSP solver used to determine the sequence in which the robot visits the groups' optimal observation positions.","marker":"[36]"},{"why":"Defines the proxemics distance scale, used to set the 0.4 m disturbance threshold that robot-group distances are checked against.","marker":"[37]"}],"fun_headline_variants":["Robot uses dipole field of personal space to find social groups","Group proxemics as magnetic field steers robot to polite spots","Vector-field model of personal space guides robot to group meetings","Robot reads social distance field to pick optimal observation points","Magnetic-dipole proxemics directs robot among social groups"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a person's social space really has the shape of the two-dimensional magnetic dipole field with the fixed constant $\\alpha=-615$, and that linearly adding those fields gives a meaningful group map whose zero-damping point is where a robot should stand; no calibration against human comfort or perception data is provided.","fun_headline_variants_meta":{"raw":{"variants":["Robot uses dipole field of personal space to find social groups","Group proxemics as magnetic field steers robot to polite spots","Vector-field model of personal space guides robot to group meetings","Robot reads social distance field to pick optimal observation points","Magnetic-dipole proxemics directs robot among social groups"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000437,"raw_usage":{"total_tokens":2256,"prompt_tokens":1015,"completion_tokens":1241,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":631,"completion_tokens_details":{"reasoning_tokens":1159}},"tokens_in":631,"tokens_out":1241,"duration_ms":9560,"temperature":1.0,"reasoning_tokens":1159,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T21:16:41.885192+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would be a perception experiment: for groups of two to four people at varied spacings and headings, compare the predicted optimal observation position from Eq. (13) with the positions human observers actually choose for a robot to approach, or with questionnaire ratings of approach comfort; if the predicted point consistently falls outside the preferred interaction region, or if small heading changes (around 10 degrees) move the predicted OOP by more than about a meter, the dipole superposition is not capturing group proxemics.","supporting_citations":[{"cited_title":"A social interaction field model accurately identifies static and dynamic social groupings,","cited_arxiv_id":null,"evidence_quote":"Supplies the social interaction field model that motivates the social-relevance term in group clustering and supports identifying static and dynamic groupings."},{"cited_title":"A New Approach for Including Social Conventions into Social Robots Navigation by Using Polygonal Triangulation and Group Asymmetric Gaussian Functions,","cited_arxiv_id":null,"evidence_quote":"Offers the existing asymmetric-Gaussian representation of individual and group zones that the proposed vector-field model is positioned against."},{"cited_title":"Space Invaders: pedestrian proxemic utility functions and trust zones for autonomous vehicle interactions,","cited_arxiv_id":null,"evidence_quote":"Contributes the idea of continuous pedestrian space functions, the conceptual basis for encoding proxemics as a field."},{"cited_title":"Safe Artificial Potential Field - novel local path planning algorithm maintaining safe distance from obstacles,","cited_arxiv_id":null,"evidence_quote":"Provides artificial potential field theory, which motivates building the guidance field by superposition of component fields."},{"cited_title":"Magne tic field diagnostics in the solar upper atmosphere,","cited_arxiv_id":null,"evidence_quote":"Supplies the planetary magnetic-field analogy that the dipole form of the individual proxemics model is drawn from."},{"cited_title":"Efficient robot motion planning using Bidirectional-Unidirectional RRT Extend function,","cited_arxiv_id":null,"evidence_quote":"Provides the classical RRT* algorithm and the Steer function that VMD-RRT* inherits and extends with the social vector field cost."},{"cited_title":"On Hierarchical Multi -UAV Dubins Traveling Salesman problem paths in a complex obst acle environment,","cited_arxiv_id":null,"evidence_quote":"Supplies the TSP solver used to determine the sequence in which the robot visits the groups' optimal observation positions."},{"cited_title":"Proxemics [and comments and replies],","cited_arxiv_id":null,"evidence_quote":"Defines the proxemics distance scale, used to set the 0.4 m disturbance threshold that robot-group distances are checked against."}],"review_version":1}