{"id":"b8e4d078-8636-46ba-918f-cef9d6f3f9c1","arxiv_id":"2607.10374","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Embedding Social Force Model dynamics and custom social costs inside nonlinear MPC yields joint human-robot trajectory prediction at 20 Hz and better social-compliance metrics than baselines in simulation.","lead":"SFM-NMPC puts the Social Force Model inside a nonlinear MPC so the robot jointly forecasts its own motion and how nearby people will react, at 20 Hz. The result is better social-distance and comfort scores than common planners in crowded simulation, which matters for service robots that must share space with people.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"SW_step metric is partially circular with the optimized J_work term, so the headline social-compliance win is not fully independent of the cost being optimized.","rationale":"The Reader correctly flags both the simulator circularity and the metric dependence; the latter is the more immediate load-bearing issue for the published claim, because SW_step is the number most prominently used to declare superiority. The paper already supplies the ablation machinery and public code, so the proposed re-evaluation is cheap and decisive. Real-robot validation remains desirable but is secondary for the present claim, which is framed entirely on simulation metrics. No change to CONDITIONAL is required; the same two fixes the Reader listed would still move the paper to ACCEPT. Confidence remains high because the equations, tables and code are fully available.","tokens_in":12664,"tokens_out":587,"duration_ms":8110,"concrete_test":"Recompute Table I and the violin/box plots of Sec. V after removing J_work from the cost (i.e., run the already-ablated SFM-NMPC(J\rho,Js\theta) variant) and report SW_step, AMD and intimate-zone intrusion on the same 30-run protocol; if SW_step advantage over MPPI-GSC/NMPC shrinks below statistical significance while AMD remains superior, the headline social win is cost-driven rather than dynamics-driven.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim rests on SFM-NMPC outperforming baselines on social-compliance metrics, especially SW_step (Table I, overall 10.70 vs next-best ~12.26) and AMD, while keeping high SR. SW_step is defined (Sec. IV-B) as the average of the social forces generated by the robot according to the Helbing SFM formulation. The same forces appear inside the optimization: J_work (Eq. 9) explicitly minimizes ||F_robot\to agent||^{2} over the horizon, and the dynamics constraints (Eqs. 3–5, 6) already embed the identical SFM. Consequently the controller is rewarded for reducing the very quantity later reported as the primary social metric. The ablation (Fig. 6) shows that adding J_work improves SW_step, confirming the direct link. AMD and proxemics (Fig. 5) are more independent, yet the paper repeatedly elevates SW_step as the decisive social-compliance advantage (Abstract, Sec. V-A, conclusions). Because the evaluation environment (HuNavSim) is itself force-based, the circularity is not diluted by an independent human model. This does not invalidate the engineering contribution or the high SR, but it weakens the claim that the method is demonstrably superior on an external social-compliance measure.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes SFM-NMPC, a nonlinear MPC local planner that embeds Helbing-style Social Force Model dynamics of surrounding agents inside the prediction model, so robot and human states are jointly propagated over a short horizon while optimizing unicycle velocity commands. A composite cost adds obstacle/costmap, goal/path, dynamic smoothness, and social terms (proxemics, social-work force magnitude, and heading long/cross penalties). The controller is implemented with Ceres, warm-started from regulated pure pursuit, and reported to run at 20 Hz. Evaluation is entirely in HuNavSim/Gazebo on open-space and mixed indoor maps (30 trials per scenario), comparing against Nav2 MPPI/DWB with Gaussian social costmaps, ORCA, SARL, pure SFM, and a non-social NMPC. The authors claim superior social compliance (especially SW_step and AMD) with high success rate, support this with Tukey HSD tests, proxemics zone occupancy, trajectory visualizations, and an ablation of social cost terms, and conclude suitability for real-world social navigation.","tokens_in":13025,"tokens_out":1668,"duration_ms":43868,"significance":"If the independent performance claims hold, this is a useful systems contribution to human-aware local planning: closed-loop multi-agent SFM prediction inside real-time NMPC is a concrete step beyond precomputed pedestrian forecasts, the reported 20 Hz rate with multi-agent dynamics is practically relevant, and the public code, multi-map protocol, statistical tests, and ablation strengthen reproducibility. Success rate in constrained mixed scenes, AMD, proxemics occupancy, and qualitative trajectory smoothness are meaningful even if SW_step is partly aligned with the objective. The advance is incremental engineering rather than a new theory of social navigation, but it is a credible addition to the MPC social-navigation literature provided the metric framing and sim-to-real caveats are tightened.","major_comments":[{"comment":"SW_step is defined (Sec. IV-B) as the average of robot-generated SFM social forces along the trajectory, using the same Helbing force structure that appears in the dynamics (Eqs. 3–5) and is explicitly minimized by J_work (Eq. 9). Table I and Sec. V-A then present the overall SW_step win (10.70 vs ~12.26 for the next-best methods) as a primary social-compliance advantage, and the ablation (Fig. 6) shows that adding J_work improves SW_step as expected. This makes the headline social-work result partly by construction rather than an external measure of comfort. AMD, proxemics (Fig. 5), success rate, path length/time, and visual interference with human paths are more independent and already favor the method in places; the manuscript should either demote SW_step to a diagnostic of the objective, report it only as an optimization residual, or add a clearly external social score (e.g., human-m","section":"Sec. IV-B, Eq. (9), Table I, Fig. 6"},{"comment":"All quantitative evidence is generated in HuNavSim, whose agents are themselves force-based, while the controller’s prediction constraints and J_work use the same SFM family (Sec. III-A, IV). Under that match, joint SFM prediction and social-work minimization are expected to look strong; the evaluation does not stress-test the weakest modeling assumption—that SFM is an adequate closed-loop predictor of how humans react to the robot over Tp = 2 s. The paper should (i) state this model-match limitation explicitly, (ii) ideally re-evaluate against at least one non-SFM pedestrian model or recorded trajectories, and (iii) avoid language that treats the current suite as sufficient confirmation of real-world social compliance. Without that, the central claim that the method is demonstrably better at social navigation remains only partially supported.","section":"Sec. III-A, Sec. IV (HuNavSim), Abstract"},{"comment":"The Abstract and Conclusions assert effectiveness “for real-world social navigation” and suitability for real deployment, yet the manuscript contains no real-robot experiments, no perception noise study, and no human-subject comfort evaluation (real-robot work is deferred to future work in Sec. VI). For a journal claim of real-world readiness this is overreach. Either add hardware or human-in-the-loop results, or rewrite the abstract/conclusion claims to match the evidence: real-time simulated social navigation under HuNavSim with improved independent metrics (SR, AMD, proxemics, trajectory quality).","section":"Abstract, Sec. VI"}],"minor_comments":[{"comment":"Cost weights α_obs, α_ρ, α_work, α_sθ, α_cross, etc., and several SFM scalars are deferred entirely to the public repository. For reproducibility in the archival version, include a table of the values used for the reported experiments (or a clear default set) in the main text or appendix.","section":"Sec. III-B"},{"comment":"Notation for agent sets and indices is dense (p^R_i, p^{A_k}_i, F_soc, etc.). A short symbol table would help readers track which quantities are optimized versus predicted.","section":"Sec. III-A–B"},{"comment":"Related work cites [25] as embedding SFM in MPC but treating humans as fixed initial predictions. A short quantitative or qualitative comparison to that class of methods (or a clear statement why it was not reimplemented) would sharpen the novelty claim of in-the-loop joint prediction.","section":"Sec. II"},{"comment":"Fig. 4 caption mentions violin plots and Tukey HSD; Fig. 6 is described as a box plot of the ablation. Ensure all significance markers and which pairs are tested are fully specified in captions so the figures stand alone.","section":"Figs. 4–6"},{"comment":"Minor wording/typos: e.g., “DWB GSC” / “MPPI GSC” spacing is inconsistent in Table I; “DWA” appears as “DWB” in places; “people’s trajectories are often precomputed even in already proposed social MPC formulations” is awkward. A careful copy-edit pass would help.","section":"Table I, Sec. I–II"},{"comment":"Prediction/control horizons (Tp = 2 s, Δts = 0.1 s) are given for NMPC, but Tc and block length B are introduced without the numerical values used in the experiments; state them explicitly.","section":"Sec. III-C, IV-A"}],"recommendation":"major_revision","confidential_remarks":"Competent IROS-style systems paper with real engineering value (joint SFM-in-the-loop NMPC at 20 Hz, code, multi-scenario stats). The main journal-level issues are metric framing (SW_step vs J_work) and sim-only overclaim under a matching force-based pedestrian model—not a broken algorithm. I would expect a revised version that demotes SW_step, foregrounds independent metrics, and tones real-world language to be acceptable; hardware results would strengthen it further but are not strictly required if claims are scoped to simulation. Fit is solid for a robotics journal with a social-navigation / MPC focus."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The real contribution here is the joint online propagation: unicycle robot plus multi-agent SFM states are rolled out together inside the same NMPC constraints at 20 Hz, plus the two heading/crossing costs that actually change the geometry of the paths. That combination is not in the cited prior work, which either freezes human forecasts or uses SFM only as a reward or pre-step. Code is public, the ablation is clean, and they beat a decent set of Nav2 baselines on success rate in the harder mixed map.\n\nWhat they do well is the engineering and the reporting. Thirty runs, Tukey tests, two maps, visual trajectories, and an incremental cost ablation. The heading terms (especially the crossing one) are the part that visibly steers the robot behind rather than in front of people; that is not just re-weighting a distance cost. AMD and the proxemics bars are independent of the optimized force term and still look favorable. Runtime claim is credible for the complexity they added.\n\nThe soft spot the stress-test flags is real but limited. SW_step is literally the quantity J_work minimizes, and HuNavSim agents are themselves force-based, so the headline social-work number is partly by construction. They lean on it too hard in the abstract and conclusions. That does not sink the paper; success rate, path quality, and the visual avoidance of forcing humans to deviate remain independent evidence. The bigger open issue is the usual one: everything is sim, and the closed-loop SFM predictor is never checked against real human reactions over the 2 s horizon. They flag real-robot work as future, which is honest.\n\nThis is for people who build local controllers for service robots and want something interpretable that runs at control rates without a learned predictor. It is not a foundational result, but it is a clean, citable engineering step. I would send it to peer review; the circularity and sim-only limits are fixable with clearer metric framing and a real-robot section, not reasons to desk-reject.","headline":"Solid real-time engineering that jointly rolls out robot and SFM humans inside NMPC; the SW_step win is partly by construction, but the rest of the evidence still holds.","tokens_in":13711,"tokens_out":508,"would_cite":true,"duration_ms":7163,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Embedding the Social Force Model inside nonlinear MPC lets a robot jointly predict human and robot motion and stay socially compliant at 20 Hz.","keywords":["social navigation","model predictive control","Social Force Model","human-aware robotics","proxemics","crowd navigation","nonlinear MPC"],"falsifier":"Run the same open-space and mixed-corridor scenarios with real pedestrians whose motion is not generated by a force model; if measured social work and intimate-space intrusion rise above the simulated baselines while success rate drops, the central claim fails.","tokens_in":13537,"feed_emoji":"🤖","tokens_out":863,"duration_ms":11091,"temperature":0.7,"pith_summary":"Robots that share space with people must do more than avoid collisions: they must anticipate how people will move and keep a comfortable distance. This paper argues that a nonlinear model-predictive controller can do that job if the Social Force Model is placed inside the prediction itself rather than treated as a fixed external forecast. At every step the optimizer rolls out both the robot and the surrounding people under the same force dynamics, then chooses velocity commands that also minimize intrusion into personal space, mutual social force, and awkward head-on or crossing headings. The resulting controller, SFM-NMPC, runs at 20 Hz and, in crowded simulated rooms and corridors, beats a range of standard planners on social-work and minimum-distance metrics while still reaching the goal reliably. The claim matters because it shows that a classical, interpretable human model can be made predictive and real-time when it is co-evolved with the robot inside the control loop.","feed_headline":"Robot predicts people with social forces inside its own planner","feed_subtitle":"SFM-NMPC jointly rolls out humans and robot, cuts social work, and still runs at 20 Hz","key_machinery":"SFM-NMPC: the Social Force Model appears both as the closed-loop dynamics that propagate every human agent inside the prediction horizon and as the source of the social-work cost term; the robot’s planned pose therefore directly shapes the predicted human reactions that the optimizer must satisfy.","core_discovery":"By writing the Social Force Model into the multi-agent dynamics of a nonlinear MPC and adding three soft social costs (proxemics, social work, and heading), the robot can jointly forecast its own trajectory and the reactive trajectories of nearby people, producing socially compliant commands that improve measured comfort metrics over strong baselines while remaining real-time at 20 Hz.","pith_inferences":["Because the simulator’s humans are themselves force-based, the reported gains may shrink once perception noise and non-force human behavior appear on a physical robot.","Co-optimizing robot and human trajectories (the paper’s stated future direction) could turn the current open-loop force prediction into a true interactive game inside the horizon.","The same embedded-dynamics pattern could be reused with any other interpretable pedestrian model (e.g., ORCA or a simple constant-velocity plus social potential) without redesigning the cost structure."],"forward_implications":["Social navigation need not rely on precomputed human trajectories or large learned predictors; a classical force model can be co-optimized with the robot in real time.","The three social cost terms (proxemics, social work, heading) are modular: any other MPC dynamics model can adopt them and expect measurable comfort gains.","Real-time execution at 20 Hz with a multi-agent force model shows that increased prediction complexity need not force a lower control rate.","Visual trajectory evidence and ablation results become standard diagnostics for claiming social compliance, not just collision-free success."],"fun_headline_variants":["MPC embeds social forces to jointly forecast people and robot","Nonlinear planner rolls out reactive humans via social force model","Social-force dynamics inside MPC yield real-time human-aware paths","Robot forecasts crowd reactions by writing social forces into MPC","Joint human-robot prediction with embedded social forces at 20 Hz"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The Social Force Model is assumed to be a good enough closed-loop predictor of how real people will react to the robot’s planned motion over the next two seconds.","fun_headline_variants_meta":{"raw":{"variants":["MPC embeds social forces to jointly forecast people and robot","Nonlinear planner rolls out reactive humans via social force model","Social-force dynamics inside MPC yield real-time human-aware paths","Robot forecasts crowd reactions by writing social forces into MPC","Joint human-robot prediction with embedded social forces at 20 Hz"]},"model":"grok-4.5","effort":"low","cost_usd":0.003496,"raw_usage":{"total_tokens":1126,"prompt_tokens":767,"num_sources_used":0,"completion_tokens":85,"cost_in_usd_ticks":34960000,"prompt_tokens_details":{"text_tokens":767,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":274,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":767,"tokens_out":85,"duration_ms":3524,"temperature":1.0,"reasoning_tokens":274,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T12:12:35.897423+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the same open-space and mixed-corridor scenarios with real pedestrians whose motion is not generated by a force model; if measured social work and intimate-space intrusion rise above the simulated baselines while success rate drops, the central claim fails.","supporting_citations":[],"review_version":1}