Pith. sign in

REVIEW 3 major objections 6 minor 35 references

In a real drugstore, 15.3% of a service robot’s turns were started by silent customer cues, and a real-time video recognizer feeding an LLM can generate proactive replies from those cues.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

About 15% of retail robot turns are nonverbal-triggered; a video multi-label recognizer plus LLM token conditioning can produce proactive responses for frequent social cues.

T0 review reviewed 2026-07-14 challenge →

load-bearing objection Solid field measurement of nonverbal-triggered turns (15.3%) and a practical cue-to-LLM pattern; the system claim is only partly backed by offline GT-track numbers and qualitative demos. the 3 major comments →

arxiv 2607.11633 v1 pith:HA2FUQTC submitted 2026-07-13 cs.RO

Breaking the 15% Barrier: A Real-World Data-Driven System for Proactive Social Robot Triggered by User Nonverbal Cues

Yuga Yano , Yuki Okafuji , Ryo Miyoshi , Sanae Yamashita , Yoshiki Ohira This is my paper
classification cs.RO
keywords social robotsnonverbal cueshuman-robot interactionretail service robotsreal-time action recognitionLLM dialogueproactive interactionmulti-label recognition
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Speech-only robot dialogue pipelines miss a real slice of store service: in a six-day teleoperated deployment, 15.3% of robot utterances were triggered by nonverbal customer acts such as approaching, waving, pointing, or showing items, not by speech. From those interactions the authors isolate nine frequent, service-relevant cues, train a real-time multi-person multi-label video recognizer, and inject the recognized cue tokens into an LLM so the robot can speak proactively without hand-crafted per-cue rules. Offline evaluation on nonverbal-triggered turns shows the system matches human teleoperator intent best on common social greetings, while rarer task directives remain harder under minimal prompting. An online prototype demonstrates live reaction to camera input. Anyone building store or public-space robots should care because audio-only systems systematically ignore openings that human operators already use.

Core claim

In a 6-day in-the-wild teleoperated deployment of a humanoid at a drugstore entrance, 15.3% of robot utterances (334 of 2,178) were initiated by users’ nonverbal behaviors rather than spoken input. From 21 observed triggering behaviors the authors select nine service-relevant cues and build a real-time multi-person, multi-label recognizer that runs online from video. They then condition LLM utterance generation on recognized nonverbal cue tokens, optionally using a vision–language model when items are shown, enabling proactive responses without hand-crafted rules; offline evaluation shows intent agreement with teleoperators especially on frequent social cues, and an online prototype reacts i

What carries the argument

A real-time multi-person multi-label nonverbal-cue recognizer (foundation video features fused with each person’s bounding-box track via attentive pooling) that emits cue tokens such as [approached] or [waved] into the LLM dialogue history, with optional VLM item description when Showed items is detected.

Load-bearing premise

The system claim rests on offline intent matching under perfect person tracks plus qualitative online demos, assuming that transfer holds once automatic detection and tracking noise are in the loop.

What would settle it

Run a controlled end-to-end online trial in the same store comparing autonomous cue-conditioned responses against teleoperator baselines on identical nonverbal-triggered moments, with automatic person detection rather than ground-truth tracks, and measure intent match plus customer engagement.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Audio-only STT–LLM–TTS pipelines systematically miss roughly 15% of customer-service robot turns that human operators already take.
  • Conditioning an LLM on a compact set of video-recognized nonverbal tokens can produce proactive greetings and engagement without per-cue rule scripts.
  • Frequent social cues (approach, wave, walk away) map reliably enough to Social & Greeting intents to support autonomous initiation.
  • Task-related directives tied to rare cues such as touching belongings still need richer store knowledge or stronger prompting.
  • Multi-person recognition above 1 fps is achievable with lightweight foundation-model features under a one-second HRI response budget.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same 15% slice implies many public-space robots (malls, stations) may under-engage if they wait only for speech.
  • Confusion between greeting and engagement invitation means timing and dialogue history, not cue identity alone, will decide which proactive move feels natural.
  • Closing the gap on task directives may require grounding the LLM in live store inventory or campaign data rather than better gesture recognition alone.
  • Offline intent agreement under ground-truth tracks likely overstates deployable performance once detector and tracker errors propagate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. This paper reports a 6-day teleoperated humanoid deployment in a drugstore and finds that 15.3% of robot utterances (334/2178) were triggered by customers’ nonverbal behaviors rather than speech, exposing a gap in audio-only STT–LLM–TTS pipelines. From 21 observed triggers the authors select nine service-relevant multi-label cues, train a real-time multi-person recognizer (V-JEPA2 features gated by box trajectories, attentive pooling, BCE with inverse-frequency weights) that reports micro-F1 0.80 under ground-truth tracks, and condition LLM utterance generation on cue tokens (with optional VLM for shown items). Offline intent-category agreement on nonverbal-triggered turns is 49.1% overall (69% on Social & Greeting) versus a 44% majority baseline; an online prototype with automatic detection/tracking is demonstrated qualitatively.

Significance. The core empirical count—that roughly one in six teleoperator turns in a real retail entrance is nonverbal-triggered—is a clear, falsifiable field measurement that is useful to HRI and service-robot dialogue design. The paper’s strengths include double-checked annotation of a non-trivial in-the-wild corpus, an explicit real-time budget (>1 fps), a compact operational taxonomy tied to observed dialogue functions (Tables I–III), and a deployable recognition architecture that avoids full VLM latency. If the transfer from oracle offline numbers to robust online proactive dialogue were better established, the work would be a solid systems contribution; even as analysis-plus-prototype it is a credible step beyond speech-only cascaded pipelines.

major comments (3)
  1. [Sec. V.B / Fig. 5] Sec. V.B and Fig. 5: the system claim that cue-conditioned LLM generation “matches teleoperator intent on frequent social cues” rests on 49.1% overall intent agreement (vs 44.0% majority baseline) under ground-truth boxes, tracks, and dialogue history. The margin is small, Task-related info & directives remain poorly recovered, and Social & Greeting vs Engagement invitation confusion is inherent in Table III. This is too thin, under oracle sensing, to underwrite the proactive-system claim; at minimum report the same intent metric with the automatic detector/tracker used online, or a controlled end-to-end user study.
  2. [Table IV / Sec. IV.B] Table IV and Sec. IV.B: all recognition precision/recall/F1 figures are computed with dataset ground-truth bounding boxes and tracking IDs. The online prototype (Sec. V.C) switches to DEIMv2-based detection/tracking, yet no multi-person multi-label metrics under automatic tracks are given. Because compounding detection/tracking error is load-bearing for the “real-time multi-person” claim, report recognition performance (or at least per-class F1) with the deployed tracker, or quantify error propagation into cue tokens.
  3. [Sec. III.A] Sec. III.A: of 32 operational hours only 1 h 42 min—described as “all interaction time where dialogue occurred”—is annotated. The 15.3% statistic and the nine-cue frequencies therefore depend on how “dialogue occurred” was defined and segmented. Clarify selection criteria, total customer approaches vs annotated segments, and whether silent or aborted approaches were excluded; without this the headline percentage’s external validity is hard to judge.
minor comments (6)
  1. [Title] Title “Breaking the 15% Barrier” overstates the result: the paper measures the barrier under teleoperation and partially addresses frequent social cues; autonomous recovery of the full 15.3% is not demonstrated.
  2. [Fig. 5] Fig. 5 intent labels are numbered 1–4 in the caption but named in the text; a single consistent legend would reduce reader load.
  3. [Sec. V.B] Sec. V.B: intent labels for generated utterances are assigned by human annotators; report inter-annotator agreement (or double-check rate) for those labels, as done for the teleoperator data.
  4. [Eq. (7)] Eq. (7): class weight w_c = 1/n_c is standard inverse frequency; state whether n_c is counted per clip, per person-instance, or per frame, and whether weights are re-normalized per batch.
  5. [Sec. V] Related work on turn-taking (VAP) is cited for timing but the prototype does not integrate it; a short note on how cue tokens interact with turn-taking would clarify scope.
  6. [References] Minor wording: “Journal of Humam-Robot Interaction” in Ref. [6] is a typo for Human.

Circularity Check

0 steps flagged

No circularity: empirical annotation counts, design choices, and standard supervised recognition/generation evaluation; nothing reduces by construction to its inputs.

full rationale

This is an applied HRI systems paper whose load-bearing claims are observational counts and engineering evaluations, not first-principles derivations. The 15.3% figure (334/2178 nonverbal-triggered utterances) is a direct annotation tally of teleoperated turns, not a quantity fitted then re-predicted. Selection of nine cues from 21 observed behaviors is an explicit design prioritization by frequency and service relevance, not a uniqueness theorem or self-definitional identity. The multi-label recognizer is trained with BCE on labeled clips and scored under 5-fold CV (micro-F1 0.80 under GT boxes); that is ordinary supervised learning, not a fitted input renamed as prediction. Offline utterance evaluation maps generated text to intent categories via independent human annotation and reports agreement (49.1% overall, 69% Social & Greeting) against teleoperator intents—again an empirical match rate, not a tautology. Self-citations to related Okafuji/Baba/etc. HRI work supply background and components but do not force the central percentage or the intent-agreement numbers. No equation chain equates a claimed prediction to its defining fit; no uniqueness is imported to forbid alternatives. Evaluation gaps (GT tracks offline vs. automatic online, weak task-directive recovery) are correctness/transfer issues, not circularity. Score 0 with empty steps is the honest finding.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The paper is empirical systems work. Load-bearing inputs are the teleoperation corpus and annotation schema, the choice of nine cues, inverse-frequency class weights, frozen V-JEPA2 features, and the assumption that teleoperator intents are the right target for autonomous generation. No new physical entities are postulated; the main invented constructs are the operational cue set and the prompt-token interface.

free parameters (4)
  • Nine selected nonverbal cue classes
    Chosen from 21 observed behaviors by frequency, service relevance, and video detectability (Sec. III.C); Touched belongings kept despite low count for task value. Selection is hand-designed and shapes all later recognition and generation results.
  • Inverse-frequency class weights w_c = 1/n_c
    BCE weights set from positive counts in the training folds (Eq. 7); directly affects minority-class learning for Pointed, Nodded, Touched belongings.
  • Clip length T and sliding inference window (8 frames)
    Temporal window for recognition and live inference is a design choice that trades motion context against latency; not derived from first principles.
  • Temporal filter for suppressing transient false positives
    Online system emits labels only after filtering (Sec. V.A); threshold/details not fully specified but affect which tokens reach the LLM.
axioms (4)
  • domain assumption Teleoperator nonverbal-triggered utterances are a valid gold standard for what an autonomous service robot should say and when.
    All analysis and intent agreement evaluation treat operator behavior as target (Secs. III, V.B); alternative policies (staff guidelines, customer preference) are not measured.
  • domain assumption Humans expect roughly one-second response delay, so recognition must run at ≥1 FPS excluding TTS.
    Used to define real-time and to exclude VLM baselines (Intro; Sec. IV, citing Shiwa et al.).
  • ad hoc to paper Observable multi-label video cues plus dialogue history suffice for LLM generation without hand-crafted per-cue rules.
    Minimal-prompt setting in Sec. V.B; task-related directives underperform, showing the axiom is only partly true.
  • domain assumption Standard multi-label BCE training on person-gated ViT features yields usable multi-person cue estimates under GT tracks.
    Architecture and loss in Sec. IV; evaluation deliberately holds detection/tracking perfect.
invented entities (2)
  • Nine-cue operational nonverbal taxonomy for retail service no independent evidence
    purpose: Compact recognition targets linking store behaviors to dialogue functions.
    Derived from this corpus rather than a prior standard label set; independent_evidence is limited to internal frequency/intent tables.
  • Nonverbal cue prompt tokens (e.g., [waved], [showed items]) for LLM conditioning no independent evidence
    purpose: Interface between the recognizer and cascaded STT–LLM–TTS dialogue without rule tables.
    Engineering construct; effectiveness measured only via intent agreement and demos in this paper.

reviewed 2026-07-14 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Breaking the 15% Barrier: A Real-World Data-Driven System for Proactive Social Robot Triggered by User Nonverbal Cues." pith.science (2026). https://pith.science/paper/HA2FUQTC

@misc{pith2026260711633,
  author       = {Pith},
  title        = {Pith review of: Breaking the 15% Barrier: A Real-World Data-Driven System for Proactive Social Robot Triggered by User Nonverbal Cues},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HA2FUQTC}},
  note         = {Machine review of arXiv:2607.11633}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Service robots in retail stores increasingly rely on cascaded speech pipelines (STT-LLM-TTS), yet many customer-robot interactions are initiated or guided by nonverbal behaviors such as approaching, waving, pointing, or showing items. This paper studies such cues in a real-world store deployment with a teleoperated humanoid robot and shows that a non-negligible portion of robot turns are triggered by nonverbal behaviors rather than spoken input, revealing a limitation of audio-only dialogue systems. In a 6-day in-the-wild deployment, 15.3\% of robot utterances were initiated by users' nonverbal behaviors rather than spoken input. Based on an analysis of observed customer behaviors, we define a set of frequent, service-relevant nonverbal cues and develop a real-time multi-person, multi-label recognizer that runs online from video. We then propose a dialogue framework that conditions LLM-based utterance generation on recognized nonverbal cue tokens, and optionally leverages a vision-language model when items are shown, enabling proactive robot responses without hand-crafted rules. We evaluate the approach offline on nonverbal-triggered turns and demonstrate an online prototype that reacts to users' nonverbal cues in real time.

Figures

Figures reproduced from arXiv: 2607.11633 by Ryo Miyoshi, Sanae Yamashita, Yoshiki Ohira, Yuga Yano, Yuki Okafuji.

Figure 1
Figure 1. Figure 1: In this study, we (a) collect dialogue data from a remote-controlled [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Data collection setup in the store. The operator viewed the rear [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Proposed architecture for user nonverbal behavior recognition [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ideal interaction generated by the proposed system [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Intent classification results. Label 1 denotes [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 9 linked inside Pith

  1. [1]

    Unified understanding of environment, task, and human for human-robot interaction in real-world environments,

    Y . Yano, A. Mizutani, Y . Fukuda, D. Kanaoka, T. Ono, and H. Tamukoh, “Unified understanding of environment, task, and human for human-robot interaction in real-world environments,” inIEEE International Conference on Robot and Human Interactive Commu- nication, pp. 224–230, 2024

    2024

  2. [2]

    A noise-robust turn-taking system for real-world dialogue robots: A field experiment,

    K. Inoue, Y . Okafuji, J. Baba, Y . Ohira, K. Hyodo, and T. Kawahara, “A noise-robust turn-taking system for real-world dialogue robots: A field experiment,” inIEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 874–879, 2025

    2025

  3. [3]

    Pseudo-eating behavior of service robot to improve the trustworthi- ness of product recommendations,

    Y . Okafuji, T. Ishikawa, K. Matsumura, J. Baba, and J. Nakanishi, “Pseudo-eating behavior of service robot to improve the trustworthi- ness of product recommendations,”Advanced Robotics, vol. 38, no. 5, pp. 343–356, 2024

    2024

  4. [4]

    Gpt-4o system card,

    A. Hurst, A. Lerer, A. P. Goucher, A. Perelman, A. Ramesh, A. Clark, A. Ostrow, A. Welihinda, A. Hayes, A. Radford,et al., “Gpt-4o system card,”arXiv preprint arXiv:2410.21276, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities,

    G. Comanici, E. Bieber, M. Schaekermann, I. Pasupat, N. Sachdeva, I. Dhillon, M. Blistein, O. Ram, D. Zhang, E. Rosen,et al., “Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities,”arXiv preprint arXiv:2507.06261, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Nonverbal cues in human–robot interac- tion: A communication studies perspective,

    J. Urakami and K. Seaborn, “Nonverbal cues in human–robot interac- tion: A communication studies perspective,”Journal of Humam-Robot Interaction, vol. 12, no. 2, 2023

    2023

  7. [7]

    V oicebench: Benchmarking llm-based voice assistants,

    Y . Chen, X. Yue, C. Zhang, X. Gao, R. T. Tan, and H. Li, “V oicebench: Benchmarking llm-based voice assistants,”arXiv preprint arXiv:2410.17196, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    Ex- V AD: Explainable fine-grained video anomaly detection based on visual-language models,

    C. Huang, Y . Shi, J. Wen, W. Wang, Y . Xu, and X. Cao, “Ex- V AD: Explainable fine-grained video anomaly detection based on visual-language models,” inProceedings of the 42nd International Conference on Machine Learning, vol. 267 ofProceedings of Machine Learning Research, pp. 25750–25761, PMLR, 13–19 Jul 2025

    2025

  9. [9]

    Streamingvlm: Real-time understanding for infinite video streams,

    R. Xu, G. Xiao, Y . Chen, L. He, K. Peng, Y . Lu, and S. Han, “Streamingvlm: Real-time understanding for infinite video streams,” arXiv preprint arXiv:2510.09608, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Videorefer suite: Advancing spatial- temporal object understanding with video llm,

    Y . Yuan, H. Zhang, W. Li, Z. Cheng, B. Zhang, L. Li, X. Li, D. Zhao, W. Zhang, Y . Zhuang,et al., “Videorefer suite: Advancing spatial- temporal object understanding with video llm,” inComputer Vision and Pattern Recognition Conference, pp. 18970–18980, 2025

    2025

  11. [11]

    How quickly should communication robots respond?,

    T. Shiwa, T. Kanda, M. Imai, H. Ishiguro, and N. Hagita, “How quickly should communication robots respond?,” inACM/IEEE International Conference on Human-Robot Interaction, pp. 153–160, 2008

    2008

  12. [12]

    What drives you to interact?: The role of user motivation for a robot in the wild,

    A. Koike, Y . Okafuji, K. Hoshimure, and J. Baba, “What drives you to interact?: The role of user motivation for a robot in the wild,” inACM/IEEE International Conference on Human-Robot Interaction, pp. 183–192, 2025

    2025

  13. [13]

    What is appropriate? on the assessment of human-robot proxemics for casual encounters in closed environments,

    N. E. Neef, S. Zabel, M. Lauckner, and S. Otto, “What is appropriate? on the assessment of human-robot proxemics for casual encounters in closed environments,”International Journal of Social Robotics, vol. 15, pp. 953–967, 2023

    2023

  14. [14]

    Applying general turn-taking models to conversational human-robot interaction,

    G. Skantze and B. Irfan, “Applying general turn-taking models to conversational human-robot interaction,” inACM/IEEE International Conference on Human-Robot Interaction, p. 859–868, 2025

    2025

  15. [15]

    From attraction to engagement: A robot-clerk col- laboration strategy for retail success,

    S. Song, T. Iwamoto, Y . Okafuji, J. Baba, J. Nakanishi, Y . Yoshikawa, and H. Ishiguro, “From attraction to engagement: A robot-clerk col- laboration strategy for retail success,”IEEE Robotics and Automation Letters, vol. 10, no. 7, pp. 6672–6679, 2025

    2025

  16. [16]

    Slowfast networks for video recognition,

    C. Feichtenhofer, H. Fan, J. Malik, and K. He, “Slowfast networks for video recognition,” inIEEE/CVF international conference on computer vision, pp. 6202–6211, 2019

    2019

  17. [17]

    Video action transformer network,

    R. Girdhar, J. Carreira, C. Doersch, and A. Zisserman, “Video action transformer network,” inIEEE/CVF conference on computer vision and pattern recognition, pp. 244–253, 2019

    2019

  18. [18]

    Tuber: Tubelet transformer for video action detection,

    J. Zhao, Y . Zhang, X. Li, H. Chen, B. Shuai, M. Xu, C. Liu, K. Kundu, Y . Xiong, D. Modolo,et al., “Tuber: Tubelet transformer for video action detection,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 13598–13607, 2022

    2022

  19. [19]

    Actionclip: A new paradigm for video action recognition,

    M. Wang, J. Xing, and Y . Liu, “Actionclip: A new paradigm for video action recognition,”arXiv preprint arXiv:2109.08472, 2021

    Pith/arXiv arXiv 2021

  20. [20]

    Sam 2: Segment anything in images and videos,

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. Rädle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V . Alwala, N. Carion, C.-Y . Wu, R. Girshick, P. Dollár, and C. Feichtenhofer, “Sam 2: Segment anything in images and videos,”arXiv preprint arXiv:2408.00714, 2024

    Pith/arXiv arXiv 2024

  21. [21]

    Interactive multimodal robot dialog using pointing gesture recogni- tion,

    S. Constantin, F. I. Eyiokur, D. Yaman, L. Bärmann, and A. Waibel, “Interactive multimodal robot dialog using pointing gesture recogni- tion,” inEuropean Computer Vision Association Workshops, pp. 640– 657, 2023

    2023

  22. [22]

    Do you follow?: A fully automated system for adaptive robot presenters,

    A. Axelsson and G. Skantze, “Do you follow?: A fully automated system for adaptive robot presenters,” inACM/IEEE International Conference on Human-Robot Interaction, pp. 102–111, 2023

    2023

  23. [23]

    When robots get chatty: Grounding multimodal human-robot conversation and collaboration,

    P. Allgeuer, H. Ali, and S. Wermter, “When robots get chatty: Grounding multimodal human-robot conversation and collaboration,” arXiv preprint arXiv:2407.00518, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    A mismatch in the human realism of face and voice produces an uncanny valley,

    W. J. Mitchell, S. Kevin A Szerszen, A. S. Lu, P. W. Schermerhorn, M. Scheutz, and K. F. MacDorman, “A mismatch in the human realism of face and voice produces an uncanny valley,”i-Perception, vol. 2, no. 1, pp. 10–12, 2011

    2011

  25. [25]

    A comprehensive review of data-driven co-speech gesture generation,

    S. Nyatsanga, T. Kucherenko, C. Ahuja, G. E. Henter, and M. Neff, “A comprehensive review of data-driven co-speech gesture generation,” arXiv preprint arXiv:2301.05339, 2023

    Pith/arXiv arXiv 2023

  26. [26]

    The iso standard for dialogue act annotation, second edition,

    H. Bunt, V . Petukhova, E. Gilmartin, C. Pelachaud, A. Fang, S. Keizer, and L. Prévot, “The iso standard for dialogue act annotation, second edition,” inInternational Conference on Language Resources and Evaluation, pp. 549–558, 2020

    2020

  27. [27]

    A survey on dialogue systems: Recent advances and new frontiers,

    H. Chen, X. Liu, D. Yin, and J. Tang, “A survey on dialogue systems: Recent advances and new frontiers,”ACM SIGKDD Explorations Newsletter, vol. 19, no. 2, pp. 25–35, 2017

    2017

  28. [28]

    Finding appropriate interaction strategies for proactive dialogue systems—an open quest,

    F. Nothdurft, S. Ultes, and W. Minker, “Finding appropriate interaction strategies for proactive dialogue systems—an open quest,” inEuropean and the 5th Nordic Symposium on Multimodal Communication, pp. 73– 80, 2015

    2015

  29. [29]

    Unraveling the thread: Understanding and addressing sequential failures in human–robot interaction,

    L. Tisserand, B. Stephenson, H. Baldauf-Quilliatre, M. Lefort, and F. Armetta, “Unraveling the thread: Understanding and addressing sequential failures in human–robot interaction,”Frontiers in Robotics and AI, vol. 11, p. 1359782, 2024

    2024

  30. [30]

    How to approach humans? strategies for social robots to initiate interaction,

    S. Satake, T. Kanda, D. F. Glas, M. Imai, H. Ishiguro, and N. Hagita, “How to approach humans? strategies for social robots to initiate interaction,” inACM/IEEE International Conference on Human-Robot Interaction, 2009

    2009

  31. [31]

    Investigating the influence of task-dependent and task-independent robot behavior on the impression of robots and the user experience,

    Y . Chamoto, Y . Okafuji, K. Matsumura, J. Baba, and J. Nakanishi, “Investigating the influence of task-dependent and task-independent robot behavior on the impression of robots and the user experience,” inIEEE International Conference on Robot and Human Interactive Communication, pp. 640–646, 2023

    2023

  32. [32]

    V-jepa 2: Self- supervised video models enable understanding, prediction and plan- ning,

    M. Assran, A. Bardes, D. Fan, Q. Garrido, R. Howes, M. Muckley, A. Rizvi, C. Roberts, K. Sinha, A. Zholus,et al., “V-jepa 2: Self- supervised video models enable understanding, prediction and plan- ning,”arXiv preprint arXiv:2506.09985, 2025

    Pith/arXiv arXiv 2025

  33. [33]

    472_deimv2-wholebody34: Lightweight human de- tection models generated on high-quality human data sets,

    K. Hyodo, “472_deimv2-wholebody34: Lightweight human de- tection models generated on high-quality human data sets,”

  34. [34]

    Available:https://github.com/PINTO0309/PINTO_ model_zoo/tree/main/472_DEIMv2-Wholebody34

  35. [35]

    Real-time object detection meets dinov3,

    S. Huang, Y . Hou, L. Liu, X. Yu, and X. Shen, “Real-time object detection meets dinov3,”arXiv preprint arXiv:2509.20787, 2025

    arXiv 2025

This paper was first reviewed by grok-4.5 on July 14, 2026.