Pith. sign in

REVIEW 3 major objections 4 minor 45 references

Foundation robot policies strong at isolated manipulation fail in collaboration: best success is 0.533 on Instructor, 0.500 on Collaborator, and 0.100 on Intruder tasks.

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 ↗

T0 review · deepseek-v4-flash

2026-08-02 08:38 UTC pith:GPN45RMC

load-bearing objection A solid diagnostic benchmark for interaction-aware VLA policies, but the headline sim2real gain is confounded by data volume and the paper needs controls, error bars, and a release before the transfer claim holds. the 3 major comments →

arxiv 2607.13056 v1 pith:GPN45RMC submitted 2026-07-05 cs.RO cs.LG

HRIBench: Benchmarking Interaction-Centric Human-Robot Collaboration

Chang Liu , Jiawei Zhang , Tao Zhang , Ye Wang , Hongyu Zhou , Qin Jin This is my paper
classification cs.RO cs.LG
keywords human-robot collaborationvision-language-action modelsbenchmarkscenario scriptsinteraction-centric evaluationtemporal coordinationintent understandingsim-to-real transfer
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 paper tries to prove a gap: robot policies with strong isolated manipulation skills are much weaker when the task involves a human partner. It introduces HRIBench, a benchmark that models collaboration as executable scenario scripts with explicit roles, timing constraints, and human motion distributions, then measures policy behavior under Instructor, Collaborator, and Intruder roles. The best evaluated policies reach a Collaboration Success Rate of 0.533 on Instructor tasks, 0.500 on Collaborator tasks, and only 0.100 on Intruder tasks, where a human disrupts execution. The paper also reports that fine-tuning on HRIBench-generated simulation data raises a physical robot's task success from 0.10 to 0.43 compared with post-training on real demonstrations alone. The significance for a reader is that evaluation of robot competence shifts from object-centric goals to coordination: who acts, when, and in response to whom.

Core claim

The paper's central claim is that interaction structure — who acts, when, in response to whom, and under what constraints — determines collaborative success more than manipulation competence. To make this measurable, the authors define 'scenario scripts': structured sequences of interaction acts, each encoding an active role, a semantic goal, temporal or causal constraints, a human motion distribution, and reward/success conditions, which are compiled into simulated episodes. The benchmark organizes 13 tasks into three roles: Instructor (intent communication), Collaborator (joint coordination), and Intruder (robustness under human intervention). Evaluating VLA policies GR00T N1.5 and pi0.5 (

What carries the argument

The central mechanism is the 'scenario script': a sequence of interaction acts, each specifying the active role, the semantic goal, temporal or causal constraints, the human motion distribution, and the reward/success conditions. These scripts are compiled into simulation environments through a four-stage pipeline (scripting, scene instantiation, motion synthesis, simulation validation), making collaboration an executable contract rather than an implicit behavior. The three interaction roles — Instructor, Collaborator, Intruder — provide a taxonomy of coordination structures, and the interaction-centric metrics (synchronization, response latency, order compliance, collision-free rate, human

Load-bearing premise

The real-world adaptation result attributes the 0.10-to-0.43 improvement to HRIBench's interaction structure, but the Sim+Real condition also has more total training data than the Real-only condition, so without a matched-data-volume baseline the specific contribution of interaction structure is not established.

What would settle it

Compare Sim+Real post-training against Real-only post-training with equal total training trajectories — for example, 40 real demonstrations versus 20 real plus 20 simulated. If the success rates become similar, the reported gain is due to data volume rather than to the interaction content of HRIBench. Conversely, if the gap persists under matched volume, the interaction-structure claim is supported.

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

If this is right

  • If the central claim is right, manipulation-centric benchmarks overstate the readiness of current robot policies for human environments; collaboration success should be measured with interaction structure built in.
  • Fine-tuning on interaction-structured generated data, even in simulation, can improve real-world collaborative performance when real demonstrations are scarce.
  • The Intruder role exposes a specific weak spot: policies can follow nominal task ordering (near-perfect order compliance) yet fail to yield, maintain clearance, or recover after intervention, indicating a missing safety-and-recovery capability.
  • Interaction-centric metrics (synchronization, responsiveness, protocol compliance, safety) can localize failures to specific coordination deficiencies rather than treating them as generic task failure.
  • HRIBench's combination of typed scripts, generated human motion, simulation filtering, and role-conditioned evaluation provides a reusable protocol for evaluating intent-aware collaboration.

Where Pith is reading between the lines

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

  • An untested implication is that the reported real-world gain may stem from extra training data volume rather than from the interaction structure of the generated episodes, because the Real-only baseline uses fewer total trajectories than Sim+Real; a matched-data comparison would be needed to attribute the gain to interaction content.
  • If scenario scripts become a standard abstraction, they could extend to other human roles (supervisor, bystander, teacher) and to multi-human teams, or to social norms beyond physical safety, such as proxemics and gaze.
  • The four-stage generation pipeline could in principle be reused to create interaction-rich training data for other robot embodiments, though this is beyond what the paper demonstrates.
  • A policy that pretrains on interaction awareness (e.g., predicting human intent or role) might be the next step the benchmark is designed to motivate.

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 / 4 minor

Summary. The paper introduces HRIBench, a diagnostic benchmark for interaction-centric human-robot collaboration. The benchmark is built on 'scenario scripts' that explicitly encode interaction roles, temporal dependencies, coordination constraints, and human motion distributions, and it defines three interaction roles — Instructor, Collaborator, and Intruder — instantiated in 13 tasks. A four-stage pipeline generates, simulates, and filters trajectories (786 of 2,000 accepted). Evaluation uses interaction-centric metrics (CSR, CT, Rsp, TSync, OC, Safety). The authors evaluate GR00T N1.5, pi0.5, and ACT under a unified protocol and find that all policies perform poorly in Intruder scenarios (CSR=0.100), with the best CSR 0.533/0.500 on Instructor/Collaborator tasks. A small real-world adaptation study on an SO-100 arm reports that HRIBench simulation data improves GR00T N1.5 physical-task success from 0.10 (Real-only) to 0.43 (Sim+Real).

Significance. If the diagnostic finding holds, HRIBench fills a genuine gap: current VLA benchmarks largely ignore interaction structure, and this paper provides a concrete, role-conditioned protocol with interpretable failure metrics. The main Table 2 result — high manipulation competence coexisting with very low Intruder CSR despite near-perfect Order Compliance — is a useful, falsifiable observation about current VLA policies. The scenario-script abstraction and the explicit modeling of human motion distributions are strengths. However, the real-world transfer claim is currently confounded, the episode-count reporting is inconsistent, and no variance or confidence intervals are given; these issues prevent the paper from establishing the benchmark's downstream value as stated. The paper also does not mention code/data release. The central diagnostic claim is plausible and likely salvageable, but the current manuscript overstates the transfer evidence.

major comments (3)
  1. [§5.3 / Table 3] The headline real-world improvement (Real-only average physical-task success 0.10 vs Sim+Real 0.43) is confounded: the two conditions differ in total training data volume and training schedule (Sim+Real adds a simulation phase plus the same 20 real demonstrations per task). The reported gain cannot be attributed to the interaction structure of HRIBench episodes without controls such as (i) Sim-only post-training, (ii) Real-matched-volume (e.g., 60 real demos per task), and (iii) simulation data matched in volume but with interaction structure ablated. The per-task results (2/10 vs 5/10, 1/10 vs 4/10, 0/10 vs 4/10) are also within the range of chance variation at n=10, so confidence intervals or a null-hypothesis test are needed. Since the abstract and conclusion use this result to claim HRIBench's value, this is a load-bearing point.
  2. [§5.1 / Table 2 / Abstract] The paper claims 'over 650 evaluation episodes' (Abstract; §3) yet §5.1 states that for each task 50 validated trajectories are used for policy adaptation and 10 held-out trajectories for evaluation. For 13 tasks this yields only 130 held-out episodes. If Table 2 is computed on the 10 held-out trajectories per task, the per-role sample sizes are small (roughly 30–50 per model), and the abstract's episode count is inconsistent. Please report exact n per role, per-task CSR, and confidence intervals or bootstrap ranges for all Table 2 metrics; with single-point metrics and no variance, the role-difficulty ordering and model comparisons cannot currently be assessed statistically.
  3. [§4 / filtering pipeline] The filtering step rejects 1,214 of 2,000 generated trajectories (786 accepted), including manual inspection for 'executability and semantic correctness.' The manuscript does not analyze whether this filtering changes the distribution of interaction patterns (e.g., which failure modes, human motion distributions, or coordination constraints survive). If the retained episodes skew toward easy or otherwise non-representative interactions, the diagnostic conclusions (especially the Intruder CSR=0.100) may not reflect the generated scenario distribution. Please report filtering statistics per task/role, the manual filtering criteria, inter-annotator agreement if applicable, and a comparison of accepted vs rejected trajectory properties.
minor comments (4)
  1. [Table 2] The 'Safety' metric is described as a compact normalized summary of CFR, HCS, CIR, and DSR, but the aggregation formula is not given in the main text. Please provide the exact definition so the safety values can be interpreted reproducibly.
  2. [References [13], [21]] Reference [13] lists 'P. Intelligence' as an author; this should be 'Physical Intelligence.' Additionally, reference [21] is also titled 'HRIBench,' which creates a naming collision with the present benchmark. Please clarify the relationship or distinguish the two benchmarks explicitly.
  3. [§5.2] The text says evaluation uses 'matched data splits, observation preprocessing, and evaluation seeds whenever supported,' but no seed values or number of seeds are reported. Please specify these details for reproducibility.
  4. [§3 / Appendix A] Key interaction metrics (TSync, Rsp, OC, CFR, HCS, CIR, DSR) are only defined in Appendix A. Since these metrics are central to the paper's contribution, a brief definition in the main text would improve readability and self-containedness.

Circularity Check

0 steps flagged

No significant circularity: benchmark diagnostics are self-contained; the real-world transfer gain is an empirical confound, not a circular derivation.

full rationale

The paper makes no equation-level derivation claims. Scenario scripts (Eq. 1) are an input representation; the evaluation metrics are defined from task outcomes and interaction events, not from the evaluated policies themselves. The central diagnostic finding (Table 2) is an empirical measurement on held-out generated episodes, and no fitted parameter is renamed as a prediction. The only self-referential element is that HRIBench generates its own simulation data and then uses that data to demonstrate its own value (Section 5.3, Table 3). This is a benchmark-construction and experimental-design concern: Sim+Real adds both HRIBench interaction data and additional training data, so the 0.10-to-0.43 physical-task gain does not isolate interaction-specific structure. However, that is an empirical confound / lack of a matched-volume control, not circularity under the rubric. No load-bearing self-citation appears; the cited work [21] with a similar name is by different authors and is not load-bearing. The Limitations section explicitly acknowledges simulation fidelity and the small scale of the SO-100 study. The claim that fine-tuning consistently improves collaborative performance is not fully demonstrated by Table 2 alone, but absence of a baseline is a completeness issue, not circular reasoning.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 1 invented entities

The ledger shows that HRIBench's central claims depend less on fitted numeric parameters and more on domain assumptions about human motion realism, the representativeness of three interaction roles, and the validity of approximate contact modeling. The only explicit free choice identified in the main text is the unstated Safety aggregation. No physical entities such as new forces or particles are introduced.

free parameters (1)
  • Safety summary aggregation = not specified (deferred to Appendix H.3)
    Table 2 reports a single 'Safety' number per model and role, stated to be a compact normalized summary of CFR, HCS, CIR, and DSR. The aggregation formula is not given in the main text, so the scalar comparison is a free choice of the authors and not independently reproducible from the preprint.
axioms (4)
  • domain assumption The three-role taxonomy (Instructor, Collaborator, Intruder) spans the coordination structures that matter for human-robot collaboration.
    The benchmark's diagnostic value depends on these three roles being representative instead of arbitrary. The paper grounds them in HRI literature (Argall, Dragan, Villani), but completeness of the taxonomy is asserted, not proven.
  • domain assumption HY-Motion-1.0-generated human skeletons, refined with labeled constraints, are a valid proxy for real human behavior during collaboration.
    All simulation episodes are built from generated human motion. If the motion prior is unrealistic, measured policy failures could be artifacts. The paper acknowledges this in Section 7 but provides no comparison against real human motion distributions.
  • domain assumption Contact and safety can be approximated by hand labels and object-proximity constraints rather than high-fidelity physical interaction.
    The safety metrics (CFR, HCS) rest on this approximation. Section 7 explicitly states that contact is approximated, meaning collision-free scores may not correspond to physical safety.
  • ad hoc to paper Manual filtering of 786 of 2,000 generated trajectories does not bias the retained interaction distribution.
    Episodes are manually inspected for executability and semantic correctness (Section 4). No inter-annotator agreement or bias analysis is reported, so the retained set could favor structures that are easy to generate rather than representative of real collaboration.
invented entities (1)
  • Scenario scripts (S = {a_1, ..., a_T}) no independent evidence
    purpose: A structured executable representation of interaction episodes, specifying roles, goals, temporal constraints, human motion distributions, and reward/success conditions.
    This is the paper's central conceptual invention. It could serve external users as a benchmark format, but no code, schema, or data release is provided in the preprint, so there is no falsifiable handle outside the paper yet.

pith-pipeline@v1.3.0-alltime-deepseek · 9515 in / 11135 out tokens · 124227 ms · 2026-08-02T08:38:15.159116+00:00 · methodology

0 comments
read the original abstract

Current vision-language-action (VLA) benchmarks primarily evaluate isolated manipulation skills while leaving human-robot interaction structure largely unmodeled. However, real-world collaboration fundamentally requires coordination under shared agency, including intent understanding, temporal synchronization, protocol adherence, and safe interaction in dynamic environments. To address this gap, we introduce HRIBench, a diagnostic benchmark for intent-aware human-robot collaboration based on executable interaction scenarios. HRIBench represents collaborative tasks as structured scenario scripts that explicitly model agent roles, temporal dependencies, coordination constraints, and human behavior distributions. Building on this abstraction, HRIBench defines three representative interaction roles: Instructor, Collaborator, and Intruder, covering intent communication, joint coordination, and robustness under human intervention. The benchmark contains 13 role-conditioned tasks with over 650 evaluation episodes generated from diverse interaction trajectories and scene variations. Beyond binary task success, HRIBench introduces interpretable interaction-centric metrics spanning synchronization, responsiveness, protocol compliance, and safety. We evaluate adapted policies based on GR00T, pi0.5, and ACT under a unified protocol. Results show that current foundation robot policies struggle substantially in collaborative settings despite strong manipulation ability, revealing major limitations in temporal coordination and intent-aware behavior. Fine-tuning on HRIBench consistently improves collaborative performance. In a real-world adaptation study, simulation data generated by HRIBench improves GR00T N1.5's physical-task success rate from 0.10 to 0.43, demonstrating the benchmark's value for advancing interaction-centric robot learning.

Figures

Figures reproduced from arXiv: 2607.13056 by Chang Liu, Hongyu Zhou, Jiawei Zhang, Qin Jin, Tao Zhang, Ye Wang.

Figure 1
Figure 1. Figure 1: HRIBench evaluates robot policies in executable human-robot interaction scenarios. Each [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: HRIBench generation pipeline. A language model creates scenario scripts, environment [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Representative real-world SO-100 tasks: gesture-conditioned delivery, collaborative re [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Representative successful and failed HRIBench rollouts across different interaction roles. [PITH_FULL_IMAGE:figures/full_fig_p008_4.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

45 extracted references · 11 linked inside Pith

  1. [1]

    Zitkovich, T

    B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahid, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In Conference on Robot Learning, pages 2165–2183. PMLR, 2023

    2023

  2. [2]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, et al. Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

    Pith/arXiv arXiv 2024

  3. [3]

    Black, N

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, et al.pi 0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    M. A. Goodrich and A. C. Schultz. Human–robot interaction: a survey.F oundations and trends® in human–computer interaction, 1(3):203–275, 2008

    2008

  5. [5]

    T. B. Sheridan. Human–robot interaction: status and challenges.Human factors, 58(4):525– 532, 2016

    2016

  6. [6]

    A. D. Dragan, K. C. Lee, and S. S. Srinivasa. Legibility and predictability of robot motion. In2013 8th ACM/IEEE International Conference on Human-Robot Interaction (HRI), pages 301–308. IEEE, 2013

    2013

  7. [7]

    Huang and B

    C.-M. Huang and B. Mutlu. Anticipatory robot control for efficient human-robot collaboration. In2016 11th ACM/IEEE international conference on human-robot interaction (HRI), pages 83–90. IEEE, 2016

    2016

  8. [8]

    K. W. Strabala, M. K. Lee, A. D. Dragan, J. L. Forlizzi, S. Srinivasa, M. Cakmak, and V . Mi- celli. Towards seamless human-robot handovers.Journal of Human-Robot Interaction, 2(1): 112–132, 2013

    2013

  9. [9]

    P. A. Lasota, T. Fong, and J. A. Shah. A survey of methods for safe human-robot interaction. F oundations and Trends® in Robotics, 5(4):261–349, 2017

    2017

  10. [10]

    Villani, F

    V . Villani, F. Pini, F. Leali, and C. Secchi. Survey on human–robot collaboration in industrial settings: Safety, intuitive interfaces and applications.Mechatronics, 55:248–266, 2018

    2018

  11. [11]

    B. D. Argall, S. Chernova, M. Veloso, and B. Browning. A survey of robot learning from demonstration.Robotics and autonomous systems, 57(5):469–483, 2009

    2009

  12. [12]

    Bjorck, F

    J. Bjorck, F. Casta ˜neda, N. Cherniadev, X. Da, R. Ding, L. Fan, Y . Fang, D. Fox, F. Hu, S. Huang, et al. Gr00t n1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734, 2025

    Pith/arXiv arXiv 2025

  13. [13]

    Intelligence, K

    P. Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, et al.pi 0.5: a vision-language-action model with open-world generaliza- tion.arXiv preprint arXiv:2504.16054, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn. Learning fine-grained bimanual manipulation with low-cost hardware.arXiv preprint arXiv:2304.13705, 2023

    Pith/arXiv arXiv 2023

  15. [15]

    C. Li, R. Zhang, J. Wong, C. Gokmen, S. Srivastava, R. Mart ´ın-Mart´ın, C. Wang, G. Levine, M. Lingelbach, J. Sun, et al. Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. InConference on Robot Learning, pages 80–93. PMLR, 2023

    2023

  16. [16]

    Nasiriany, A

    S. Nasiriany, A. Maddukuri, L. Zhang, A. Parikh, A. Lo, A. Joshi, A. Mandlekar, and Y . Zhu. Robocasa: Large-scale simulation of everyday tasks for generalist robots.arXiv preprint arXiv:2406.02523, 2024. 9

    Pith/arXiv arXiv 2024

  17. [17]

    James, Z

    S. James, Z. Ma, D. R. Arrojo, and A. J. Davison. Rlbench: The robot learning benchmark & learning environment.IEEE Robotics and Automation Letters, 5(2):3019–3026, 2020

    2020

  18. [18]

    O. Mees, L. Hermann, E. Rosete-Beas, and W. Burgard. Calvin: A benchmark for language- conditioned policy learning for long-horizon robot manipulation tasks.IEEE Robotics and Automation Letters, 7(3):7327–7334, 2022

    2022

  19. [19]

    T. Mu, Z. Ling, F. Xiang, D. Yang, X. Li, S. Tao, Z. Huang, Z. Jia, and H. Su. Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations.arXiv preprint arXiv:2107.14483, 2021

    Pith/arXiv arXiv 2021

  20. [20]

    T. Yu, D. Quillen, Z. He, R. Julian, K. Hausman, C. Finn, and S. Levine. Meta-world: A benchmark and evaluation for multi-task and meta reinforcement learning. InConference on robot learning, pages 1094–1100. PMLR, 2020

    2020

  21. [21]

    Z. Shi, E. Zhao, N. Dennler, J. Wang, X. Xu, K. Shrestha, M. Fu, D. Seita, and M. Matari ´c. Hribench: Benchmarking vision-language models for real-time human perception in human- robot interaction.arXiv preprint arXiv:2506.20566, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    M. Jung, Q. Yang, L. von Seelstrang, D. Strazdas, S. Wachsmuth, and A. Al-Hamadi. Semiac: A two-site human-robot collaboration dataset for exploring socially-enriched models for im- plicit action coordination. InCompanion Proceedings of the 21st ACM/IEEE International Conference on Human-Robot Interaction, pages 1028–1033, 2026

    2026

  23. [23]

    Saupp ´e and B

    A. Saupp ´e and B. Mutlu. Robot deictics: How gesture and context shape referential commu- nication. InProceedings of the 2014 ACM/IEEE international conference on Human-robot interaction, pages 342–349, 2014

    2014

  24. [24]

    V . Bo, A. Garrell, and A. Sanfeliu. Fast or accurate? how intention-recognition models shape human perception of a mobile robot. InCompanion Proceedings of the 21st ACM/IEEE Inter- national Conference on Human-Robot Interaction, pages 502–506, 2026

    2026

  25. [25]

    H. Wang, K. Winkle, and G. Castellano. Anticipating safety and security risks of embodied ai: A multidimensional risk framework for foundation models in hri. InCompanion Proceedings of the 21st ACM/IEEE International Conference on Human-Robot Interaction, pages 1039– 1044, 2026

    2026

  26. [26]

    E. J. Bejarano Sepulveda, V . Bo, A. Sanfeliu, and A. Garrell. Perception–awareness–decision: Socially-aware robot navigation and interaction. InCompanion Proceedings of the 21st ACM/IEEE International Conference on Human-Robot Interaction, pages 258–262, 2026

    2026

  27. [27]

    Hoffman and C

    G. Hoffman and C. Breazeal. Effects of anticipatory action on human-robot teamwork ef- ficiency, fluency, and perception of team. InProceedings of the ACM/IEEE international conference on Human-robot interaction, pages 1–8, 2007

    2007

  28. [28]

    Cakmak, S

    M. Cakmak, S. S. Srinivasa, M. K. Lee, S. Kiesler, and J. Forlizzi. Using spatial and temporal contrast for fluent robot-human hand-overs. InProceedings of the 6th international conference on Human-robot interaction, pages 489–496, 2011

    2011

  29. [29]

    C. Wang, B. Du, J. Xu, P. Li, D. Guo, and H. Liu. Demonstrating humanthor: a simulation platform and benchmark for human-robot collaboration in a shared workspace.arXiv preprint arXiv:2406.06498, 2024

    Pith/arXiv arXiv 2024

  30. [30]

    X. Puig, E. Undersander, A. Szot, M. Dallaire Cote, T.-Y . Yang, R. Partsey, R. Desai, A. Clegg, M. Hlavac, S. Y . Min, et al. Habitat 3.0: A co-habitat for humans, avatars, and robots. In International Conference on Learning Representations, volume 2024, pages 15306–15336, 2024. 10

    2024

  31. [31]

    Chang, G

    M. Chang, G. Chhablani, A. Clegg, M. Dallaire Cote, R. Desai, M. Hlavac, V . Karashchuk, J. Krantz, R. Mottaghi, P. Parashar, et al. Partnr: A benchmark for planning and reasoning in embodied multi-agent tasks. InInternational Conference on Learning Representations, volume 2025, pages 65205–65268, 2025

    2025

  32. [32]

    Y .-W. Chao, C. Paxton, Y . Xiang, W. Yang, B. Sundaralingam, T. Chen, A. Murali, M. Cakmak, and D. Fox. Handoversim: A simulation framework and benchmark for human-to-robot object handovers. In2022 International conference on robotics and automation (ICRA), pages 6941–

  33. [33]

    Baraglia, M

    J. Baraglia, M. Cakmak, Y . Nagai, R. Rao, and M. Asada. Initiative in robot assistance during collaborative task execution. In2016 11th ACM/IEEE international conference on human- robot interaction (HRI), pages 67–74. IEEE, 2016

    2016

  34. [34]

    Y . Wang, Z. Xian, F. Chen, T.-H. Wang, Y . Wang, K. Fragkiadaki, Z. Erickson, D. Held, and C. Gan. Robogen: Towards unleashing infinite data for automated robot learning via generative simulation.arXiv preprint arXiv:2311.01455, 2023

    Pith/arXiv arXiv 2023

  35. [35]

    Y . J. Ma, W. Liang, G. Wang, D.-A. Huang, O. Bastani, D. Jayaraman, Y . Zhu, J. Fan, et al. Eureka: Human-level reward design via coding large language models. InInternational con- ference on learning Representations, volume 2024, pages 26516–26560, 2024

    2024

  36. [36]

    Zhang and H

    B. Zhang and H. Soh. Large language models as zero-shot human models for human-robot in- teraction. In2023 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 7961–7968. IEEE, 2023

    2023

  37. [37]

    C. Guo, S. Zou, X. Zuo, S. Wang, W. Ji, X. Li, and L. Cheng. Generating diverse and natural 3d human motions from text. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5152–5161, 2022

    2022

  38. [38]

    Petrovich, M

    M. Petrovich, M. J. Black, and G. Varol. Temos: Generating diverse human motions from textual descriptions. InEuropean conference on computer vision, pages 480–497. Springer, 2022

    2022

  39. [39]

    Tevet, S

    G. Tevet, S. Raab, B. Gordon, Y . Shafir, D. Cohen-Or, and A. H. Bermano. Human motion diffusion model.arXiv preprint arXiv:2209.14916, 2022

    Pith/arXiv arXiv 2022

  40. [40]

    C. Guo, Y . Mu, M. G. Javed, S. Wang, and L. Cheng. Momask: Generative masked modeling of 3d human motions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900–1910, 2024

    1900

  41. [41]

    Batool, I

    F. Batool, I. Zhura, V . Serpiva, R. A. Khan, I. Valuev, I. Tokmurziyev, and D. Tsetserukou. Humandiffusion: A vision-based diffusion trajectory planner with human-conditioned goals for search and rescue uav. InCompanion Proceedings of the 21st ACM/IEEE International Conference on Human-Robot Interaction, pages 1023–1027, 2026

    2026

  42. [42]

    Zhang, Z

    M. Zhang, Z. Cai, L. Pan, F. Hong, X. Guo, L. Yang, and Z. Liu. Motiondiffuse: Text-driven human motion generation with diffusion model.IEEE transactions on pattern analysis and machine intelligence, 46(6):4115–4128, 2024

    2024

  43. [43]

    Zhang, Y

    J. Zhang, Y . Zhang, X. Cun, Y . Zhang, H. Zhao, H. Lu, X. Shen, and Y . Shan. Generating human motion from textual descriptions with discrete representations. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14730–14740, 2023

    2023

  44. [44]

    Mainprice and D

    J. Mainprice and D. Berenson. Human-robot collaborative manipulation planning using early prediction of human motion. In2013 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 299–306. IEEE, 2013

    2013

  45. [45]

    Y . Wen, Q. Shuai, D. Kang, J. Li, C. Wen, Y . Qian, N. Jiao, C. Chen, W. Chen, Y . Wang, et al. Hy-motion 1.0: Scaling flow matching models for text-to-motion generation.arXiv preprint arXiv:2512.23464, 2025. 11

    arXiv 2025