{"id":"7859c124-9471-4bf3-b1de-e724f322fa51","arxiv_id":"2607.01287","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A VLM-based perceptual module plus MPPI controller lets robots accompany dynamically changing human groups, yielding 15% higher success rate and 25% lower collisions than baselines in five scenarios plus positive user-study ratings for naturalness.","lead":"The paper proposes using vision-language models to let robots infer good positions and distances when following human groups whose shapes change, then feeds that into an MPPI controller for safe motion. A generalist reader might care because it shows one concrete way large language models could make robots behave less awkwardly around people in real settings like events or care facilities.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"VLM position inference accuracy is unmeasured; overall metrics cannot isolate its contribution","rationale":"The reader's weakest assumption directly identifies the same unverified inference step. Because the full manuscript was not supplied to the first reader, the current pass confirms that the same gap remains the load-bearing point even after the abstract-only limitation is removed. No other internal inconsistency (e.g., in the MPPI formulation) appears more critical than this missing isolation of the VLM component.","tokens_in":1696,"tokens_out":372,"duration_ms":10750,"concrete_test":"From the methods section, extract the exact VLM prompt template and the format of the visual representation fed to it. Run the VLM on 50 held-out annotated group-formation images (with ground-truth target positions and distances) and compute mean Euclidean error plus fraction of outputs falling outside the claimed social-distance band; if mean error exceeds the tolerance used in the MPPI cost function, the inference reliability assumption fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires that the VLM, given perceptual-module visual representations, produces reliable companion positions, social distances, and dynamic group targets that the MPPI controller can then track safely. The abstract presents this inference step as the novel mechanism, yet reports only aggregate success/collision rates and a user study. No separate metric (e.g., position error, prompt consistency, or failure cases of the VLM output) is supplied, so it is impossible to tell whether the reported 15 % / 25 % gains arise from the VLM reasoning or from the MPPI safety layer, scene-specific tuning, or baseline weaknesses. If VLM outputs are frequently off by more than the social-distance tolerance, the headline improvements are not attributable to the stated core contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes an adaptive group-accompaniment method for social robots that detects group members, uses a perceptual module to generate visual representations of the interaction space, feeds these to a Vision-Language Model (VLM) to infer companion positions, social distances, and group dynamics, and integrates the VLM output with a Model Predictive Path Integral (MPPI) controller for stability and safety. It claims that experiments across five scenarios demonstrate a 15% improvement in success rate and 25% reduction in collision rate versus baselines, with a user study indicating that the behaviors are perceived as natural and socially appropriate.","tokens_in":1848,"tokens_out":472,"duration_ms":27726,"significance":"If the central claim holds after the VLM contribution is isolated and the experimental protocol is fully documented, the work would offer a concrete demonstration of combining semantic VLM reasoning with receding-horizon control for dynamic social navigation; this could inform future designs that move beyond fixed-formation assumptions in group-following tasks.","major_comments":[{"comment":"Abstract, results paragraph: the stated 15% success-rate improvement and 25% collision-rate reduction are presented without any description of the five scenarios, the baseline methods, the precise VLM prompting or fine-tuning procedure, error bars, or statistical tests; consequently the numerical claims cannot be verified and the data cannot be shown to support the headline result.","section":"Abstract"},{"comment":"Method description (abstract): the central mechanism asserts that the VLM, given perceptual-module visual representations, produces reliable companion positions, social distances, and dynamic targets that the MPPI controller can track; yet no isolated metric (position error, prompt-consistency score, or VLM failure-case analysis) is supplied, so aggregate success/collision figures cannot attribute gains to the VLM inference step rather than to the MPPI safety layer or baseline weaknesses.","section":"Method"}],"minor_comments":[{"comment":"The abstract would be clearer if it briefly indicated the output format of the perceptual module (e.g., image patches, bounding-box overlays, or scene graphs) before describing VLM input.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback highlighting the need for greater clarity in the abstract and stronger isolation of the VLM contribution. We address each major comment below.","responses":[{"response":"The abstract is intentionally concise, but the five scenarios, baseline methods, VLM prompting procedure, error bars, and statistical tests are fully documented in the Experimental Evaluation section. We will revise the abstract to briefly name the scenarios and baselines while directing readers to the full details, error bars, and significance tests in the body. This improves verifiability from the abstract without violating length limits.","revision_made":"partial","referee_comment":"[Abstract] Abstract, results paragraph: the stated 15% success-rate improvement and 25% collision-rate reduction are presented without any description of the five scenarios, the baseline methods, the precise VLM prompting or fine-tuning procedure, error bars, or statistical tests; consequently the numerical claims cannot be verified and the data cannot be shown to support the headline result."},{"response":"The reported gains are measured against baselines that omit the VLM perceptual module, so the performance delta is attributable to the addition of VLM reasoning within the integrated pipeline. To strengthen attribution, we will add a dedicated VLM evaluation subsection reporting position error and prompt-consistency metrics on the perceptual outputs.","revision_made":"yes","referee_comment":"[Method] Method description (abstract): the central mechanism asserts that the VLM, given perceptual-module visual representations, produces reliable companion positions, social distances, and dynamic targets that the MPPI controller can track; yet no isolated metric (position error, prompt-consistency score, or VLM failure-case analysis) is supplied, so aggregate success/collision figures cannot attribute gains to the VLM inference step rather than to the MPPI safety layer or baseline weaknesses."}],"tokens_in":1375,"tokens_out":403,"duration_ms":23405,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper integrates a perceptual module that feeds visual group representations into a VLM to pick companion positions and social distances, then hands those targets to an MPPI controller for safe motion. That specific stack for non-fixed formations is the concrete step it takes.\n\nIt handles a real practical problem: groups change shape, and fixed-formation methods break. The pipeline description is straightforward and the choice of MPPI for stability is reasonable given the safety needs.\n\nThe reported 15 % success gain and 25 % collision drop come from five scenarios plus a user study on perceived naturalness. Those numbers are presented as evidence that the method works.\n\nThe main weakness is that nothing isolates the VLM step. There is no metric for how accurately the VLM predicts positions or distances, no failure cases for its outputs, and no ablation that turns the VLM off. The headline improvements could come from the controller, the baselines chosen, or scene-specific tuning instead. The abstract also gives no error bars, no statistical tests, and no explicit comparison to earlier VLM-robot papers.\n\nThis is for people already working on group navigation or VLM applications in robotics. They might pick up the perceptual-module idea or the MPPI pairing, but they will not get a clear answer on whether the VLM reasoning is reliable.\n\nSend it to review. The integration is worth a closer look and referees can request the missing VLM-specific measurements, but the current evidence does not yet pin the gains on the stated core mechanism.","headline":"VLM position inference is the claimed novelty but the experiments report only aggregate success and collision numbers with no separate check on VLM output quality.","tokens_in":2357,"tokens_out":381,"would_cite":false,"duration_ms":23833,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Robots use vision-language models to adapt positions while following groups whose formations change over time.","keywords":["group following","social robots","vision-language models","dynamic formations","adaptive companionship","model predictive control"],"falsifier":"A controlled test in which the robot is placed in a previously unseen group formation and the measured success rate falls below the reported baseline.","tokens_in":2591,"feed_emoji":"🤖","tokens_out":582,"duration_ms":18481,"temperature":0.7,"pith_summary":"The paper proposes a method that lets a robot accompany a group of people even when the group members shift their relative positions. A perceptual module first creates visual representations of the group's space, which a vision-language model then uses to reason about suitable companion locations, social distances, and overall dynamics. These inferences feed into a Model Predictive Path Integral controller that generates safe, stable robot motion. Tests across five scenarios report higher success rates and fewer collisions than earlier methods, while a user study finds the resulting behaviors appear natural. The work targets the practical problem that fixed-formation techniques break down once real human groups begin to move fluidly.","feed_headline":"Vision-language models adapt robot positions for changing human groups","feed_subtitle":"Tests across five scenarios report 15 percent higher success and 25 percent fewer collisions than prior methods","key_machinery":"Vision-language model inference of companion positions and group dynamics from perceptual visual representations of the interaction space, integrated with an MPPI controller for motion generation.","core_discovery":"The central claim is that combining visual representations of group interaction space with a vision-language model's semantic reasoning, then feeding the output to a Model Predictive Path Integral controller, produces stable and socially appropriate accompaniment even as group formations change dynamically.","pith_inferences":["The same visual-reasoning loop could be applied to other social navigation tasks that require understanding changing spatial relations among people.","Real-world deployment would require checking how well the model handles partial occlusions or rapid group splits that were not emphasized in the five scenarios.","If the VLM component generalizes, it may reduce the need for hand-crafted rules about social distance in future robot navigation systems."],"forward_implications":["The approach yields a 15 percent higher success rate than baseline methods across the tested scenarios.","Collision rates decrease by 25 percent relative to the same baselines.","User evaluations rate the produced companionship behaviors as natural and socially appropriate.","The combination of perceptual module, VLM, and MPPI controller maintains stability and safety during motion."],"fun_headline_variants":["VLMs infer companion positions for group-following robots","Vision models with MPPI handle dynamic group formations","Adaptive companionship via VLM and path integral controller","VLM semantic reasoning for stable robot group accompaniment"],"cache_read_input_tokens":64,"weakest_assumption_plain":"The vision-language model can reliably interpret visual group representations to select appropriate positions and distances.","fun_headline_variants_meta":{"raw":{"variants":["VLMs infer companion positions for group-following robots","Vision models with MPPI handle dynamic group formations","Adaptive companionship via VLM and path integral controller","VLM semantic reasoning for stable robot group accompaniment"]},"model":"grok-4.3","cost_usd":0.006579,"raw_usage":{"total_tokens":3034,"prompt_tokens":590,"num_sources_used":0,"completion_tokens":58,"cost_in_usd_ticks":65787000,"prompt_tokens_details":{"text_tokens":590,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2386,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":590,"tokens_out":58,"duration_ms":17227,"temperature":1.0,"reasoning_tokens":2386,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-03T20:35:51.039653+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A controlled test in which the robot is placed in a previously unseen group formation and the measured success rate falls below the reported baseline.","supporting_citations":[],"review_version":1}