Pith. sign in

REVIEW 3 major objections 4 minor 29 references

CARMA: Context-Aware Situational Grounding of Human-Robot Group Interactions by Combining Vision-Language Models with Object and Action Recognition

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper shows that a robot system can produce grounded actor-action-object triplets for group interactions, with the best configuration reaching a 0.72 task success rate in sorting, pouring, and handover experiments.

desk verdict A useful VLM-based triplet grounding system with a thin evaluation and an unmeasured person-ID assumption that undercuts the reliability claim. read the letter →

arxiv 2506.20373 v1 pith:HV6FXFV3 submitted 2025-06-25 cs.RO cs.AIcs.HC

Joerg Deigmoeller , Stephan Hasler , Nakul Agarwal , Daniel Tanneberg , Anna Belardinelli , Reza Ghoddoosian , Chao Wang , Felix Ocker
show 3 more authors
Fan Zhang Behzad Dariush Michael Gienger
This is my paper · ORCID
classification cs.ROcs.AIcs.HC
keywords situationalgroundingactor-action-objecttriplethuman-robotgroupinteractionvision-languagemodelinstance-levelperceptionactionrecognitionobjectdetectiontasksuccessrate
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a robot can move beyond recognizing classes of objects and actions to tracking specific physical instances—this particular person, this particular cup—and representing each interaction as a grounded actor-action-object triplet. The proposed system, CARMA, combines an object detector, a person tracker, a low-level action detector, a memory of past detections, and a vision-language model that reasons over cropped images of the current person and known objects. In three tabletop experiments (sorting fruit, pouring, handover) with one or two humans and optionally a robot, the best configuration—triggering the reasoning step only when the action detector signals a new action, without feeding in the previous triplet—achieves a Task Success Rate of 0.72 across all settings. If this holds, it gives collaborative robots a structured, episodic representation of who did what with which object, useful for intention inference, ownership resolution, and memory-based decisions in shared spaces.

What carries the argument

The load-bearing structure is the actor-action-object triplet plus the memory that supplies stable instance identities. The Action Reasoner—a vision-language model guided by a fixed prompt—receives four recent cropped person images, cropped object-instance images from memory, and optionally the previous triplet or an action label; a low-level action detector decides when to trigger this reasoning by detecting action-label changes. The memory, implemented as a database of detected instances with unique IDs, cropped images, 3D locations, and timestamps, is what turns repeated class-level detections into re-identifiable individuals. The system's design claim is that the trigger signal acts as an event filter and that grounding the VLM in real detected instances keeps its output anchored to the actual scene.

What would settle it

Run the sorting task with two similarly dressed people who cross paths mid-task, manually annotate identity switches from the tracker, and compare TSR on clips with and without switches. If clips with a single detected ID switch do not show a clear drop in actor attribution accuracy, the system can tolerate tracker failures; if they do, the claimed reliable grounding is contingent on perfect tracking.

Watch

Extended reading notes

Core claim

The central claim is that instance-level situational grounding for human-robot group interaction can be achieved by coupling a memory of detected objects and tracked persons with a vision-language model, rather than by asking a VLM to parse raw video continuously. CARMA outputs triplets of the form actor–action–object (with an optional receiving actor or supporting object), where each actor and object is a unique physical instance stored in memory. The Action Reasoner is prompted with cropped person images from the tracker, cropped images of all known objects, and optionally the previous triplet or an action label from the low-level Action Detector; the detector also supplies a trigger when a new action begins. The paper reports that the trigger-only configuration outperforms continuous sampling and also beats configurations that carry the previous triplet or the raw action label, with the 0.72 aggregate TSR. The interpretation offered is that event-driven, instance-grounded prompting gives a VLM enough structure to avoid class-level confusion and spurious detections.

Load-bearing premise

The whole pipeline depends on the body pose tracker always giving every person the same unique ID across time and camera views; if two people swap IDs, every later actor-action-object triplet can attribute the action to the wrong person.

Editorial extensions

If this is right

  • Event-driven triggering beats continuous analysis: configurations using the action trigger outperform discrete sampling by about 10 percentage points overall, so perception pipelines for collaboration can reason only when low-level detectors signal a change.
  • Feeding the previous triplet back into the VLM hurts rather than helps, suggesting temporal context in prompts can bias the reasoner toward stale or incorrect assignments.
  • Giving the VLM raw action labels from the detector degrades performance most, implying the VLM over-relies on noisy low-level labels; downstream systems should treat such labels as triggers, not facts.
  • Instance-level grounding resolves ambiguities that class-level recognition cannot—ownership of an item, missing ingredients, and parallel actions by multiple actors become explicit in the triplet representation.
  • Performance in the two-person-plus-robot sorting condition stays comparable to simpler setups, indicating the approach may scale to multi-party scenes when cropping confines attention.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If person tracking is the true bottleneck, then a testable extension is to plug in a different tracker or add appearance-based re-identification; the same framework should then carry the claimed benefits without relying on the current tracker's ID stability.
  • The memory of object instances opens the door to object permanence across occlusions and long horizons—an implication the paper gestures at but does not evaluate; a longer-duration study with objects leaving and re-entering the scene would test it.
  • The failure of the previous-triplet input suggests a possible improvement the paper does not pursue: reset or weight temporal context when the action detector's trigger indicates a boundary, so the VLM is not anchored to an obsolete episode.
  • Because the system only reasons on action-trigger events, it should have lower latency and compute than continuous VLM video analysis; measuring energy or reaction-time savings in handover tasks would quantify a benefit that is only implicit here.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces CARMA, a system for situational grounding in human-robot group interactions. It combines a memory module, a 3D object detector, a body-pose-based person tracker, a low-level action detector (I3D), and a VLM-based action reasoner (GPT-4o) that receives cropped person images, object instance crops, and optional action labels/triggers to produce uniquely grounded actor-action-object triplets. The system is evaluated on three tabletop tasks (sorting fruits, pouring, handover) under up to four settings (1P, 2P, 1P+R, 2P+R), with 8 settings total and 3 runs each, yielding 24 recordings. Four system configurations are compared using a Task Success Rate (TSR) metric; the best configuration (action trigger without previous triplet) achieves an overall TSR of 0.72. The paper claims the system can reliably generate accurate triplet representations suitable for spatiotemporal reasoning and situated decision-making.

Significance. If the empirical claims hold, CARMA addresses a real gap in HRI: moving from class-level VLM understanding to instance-level, multi-actor situational grounding. The architecture is modular, combining off-the-shelf components in a sensible way, and the authors release their dataset and code (GitHub link). The evaluation across four configurations provides a useful internal ablation of the action trigger, the action label, and the previous-triplet context. However, the validation is currently too thin to support the headline claim of reliable triplet generation, and the paper does not demonstrate that the grounding components themselves provide a measurable benefit over a VLM baseline without those components.

major comments (3)
  1. [Section IV.A, Table I] The empirical basis for the central claim consists of only 24 recordings (8 settings x 3 runs). No error bars, confidence intervals, or statistical tests are reported. The claim that configurations 2 and 3 outperform others by approximately 10% and the ``exception'' in pouring of about 3% cannot be distinguished from sampling noise at this sample size. The authors should provide per-trial results, variance estimates, and, if possible, paired significance tests across configurations, or at least clearly report the qualitative patterns without overstating their reliability.
  2. [Section III, Table II] The person tracker's consistency over time is asserted but never measured. Table II explicitly excludes actors from scoring because they are ``always correctly assigned,'' yet no ID-stability metric (e.g., IDSW, IDF1, MOTA) is reported. In the 2P and 2P+R settings, two people are in view and interact (handover, crossing), which are known conditions for ID switches in body-pose trackers. A single ID switch or merge would corrupt the actor field and, consequently, the action and object attribution for all subsequent triplets. Since TSR explicitly includes ``correct actor-role attribution,'' excluding actors from the scoring inflates the reported TSR and removes the system's central failure mode from the evaluation. The authors should either report ID-stability metrics for the person tracker or include actor-assignment errors in the TSR computation.
  3. [Section IV.C, Section V] The paper claims that ``structured, grounded perception ... provides concrete functional benefits for collaborative autonomy,'' but the experiments only compare four internal configurations that all share the CARMA grounding elements (object crops, cropped person images, memory). There is no comparison to a baseline without these grounding components, such as a VLM prompted with the full camera frame and no object instance crops. As a result, the experiments cannot attribute the reported performance to the CARMA architecture itself. Adding such a baseline (or an ablation without object crops) would be needed to support the significance claim.
minor comments (4)
  1. [Table I] The column headers contain typos: ``1P-R'' appears in the Pouring and Handover sections and likely should read ``1P+R,'' and the Handover column ``I'' is ambiguous (probably ``1P''). Please correct these labels.
  2. [References] Reference [5] and reference [12] are duplicate entries for the same paper (``VLM See, Robot Do''). In addition, reference [19] (GPT-4 Technical Report) incorrectly lists arXiv:2405.21075, and reference [20] repeats the DOI from [19]; these should be corrected.
  3. [Section IV.A] The TSR metric is described only at a high level. The paper should specify how triplets are matched to ground truth: what constitutes a partial match (e.g., correct action but wrong object), how temporal ordering is scored, and how the ground-truth annotations were produced (number of annotators, annotation format, inter-annotator agreement). This detail is necessary for reproducibility.
  4. [Section V] The conclusion states that the event-driven approach ``reduces latencies, which is vital in time-sensitive interactions,'' but no latency measurements are reported in Section IV. Either provide quantitative latency data or soften this claim to a qualitative expectation.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: CARMA's triplet outputs are scored against external ground-truth sequences, and the few self-citations appear only as related work.

full rationale

The derivation chain in CARMA is modular: object instances, person IDs, action labels, and VLM reasoning are independent inputs combined by the Action Reasoner, and the final triplets are compared with manually created ground-truth annotations. No fitted parameter is renamed as a prediction, and the 'previous triplet' prompt input is temporal context rather than the quantity being predicted. The self-citations (LaMI, Attentive Support, memory-net action patterns) are used as background or inspiration and are not load-bearing: none of them is invoked to establish the correctness of the reported TSR. The one concern flagged in the manuscript—Table II excludes actors from scoring 'as they are always correctly assigned'—is an unsupported empirical assumption about person-tracker ID stability, and a real evaluation gap in a multi-person setting, but it is not a circular reduction: the paper does not define actor correctness in terms of its own outputs, nor does it fit the actor assignment to the target triplets. Therefore no specific equation or definition reduces a claimed result to its own input, and the correct finding is no significant circularity.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

CARMA introduces no new physical entities or forces. The main dependencies are on standard vision components and a proprietary VLM. The free parameters are limited to design choices in the action-trigger pipeline, and the key axioms concern the reliability of person tracking and object segmentation in uncontrolled multi-person scenes.

free parameters (2)
  • Action Detector four-frame window size = 4 frames
    The Action Detector groups cropped person images into four-frame sequences; this number is chosen without ablation or justification (Section III, Action Detector).
  • Action trigger threshold
    The condition for triggering the Action Reasoner when an action label changes is not specified numerically; it is a free design choice that affects timing and performance (Section III).
assumptions (4)
  • domain assumption 3D point cloud clustering faithfully segments all relevant objects in the workspace without merging or missing instances.
    In the Object Detector paragraph, the system relies on point cloud clustering to produce clean object candidates; failures in segmentation directly propagate to triplet object assignment.
  • domain assumption The body pose tracker assigns and maintains a unique ID per person over time, even with multiple persons and occlusions.
    The Person Tracker paragraph states unique IDs are assigned to each detected individual; no ID-switch statistics are provided, yet actor attribution in every triplet depends on this.
  • domain assumption The VLM (GPT-4o) can reliably map cropped object images plus captions to physical instances and infer actions and object relations from the provided prompt.
    The Action Reasoner uses a prompt with cropped images and captions as the grounding mechanism; the paper provides no independent evaluation of this mapping beyond the final TSR.
  • domain assumption The I3D-based action detector provides useful prior and trigger information in real time.
    The Action Detector paragraph assumes four-frame sequences yield action labels with enough accuracy to drive triggering; configuration 4 results show these labels can actively mislead, indicating this assumption is fragile.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CARMA: Context-Aware Situational Grounding of Human-Robot Group Interactions by Combining Vision-Language Models with Object and Action Recognition." pith.science (2026). https://pith.science/paper/HV6FXFV3

@misc{pith2026250620373,
  author       = {Pith},
  title        = {Pith review of: CARMA: Context-Aware Situational Grounding of Human-Robot Group Interactions by Combining Vision-Language Models with Object and Action Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HV6FXFV3}},
  note         = {Machine review of arXiv:2506.20373}
}
read the original abstract

We introduce CARMA, a system for situational grounding in human-robot group interactions. Effective collaboration in such group settings requires situational awareness based on a consistent representation of present persons and objects coupled with an episodic abstraction of events regarding actors and manipulated objects. This calls for a clear and consistent assignment of instances, ensuring that robots correctly recognize and track actors, objects, and their interactions over time. To achieve this, CARMA uniquely identifies physical instances of such entities in the real world and organizes them into grounded triplets of actors, objects, and actions. To validate our approach, we conducted three experiments, where multiple humans and a robot interact: collaborative pouring, handovers, and sorting. These scenarios allow the assessment of the system's capabilities as to role distinction, multi-actor awareness, and consistent instance identification. Our experiments demonstrate that the system can reliably generate accurate actor-action-object triplets, providing a structured and robust foundation for applications requiring spatiotemporal reasoning and situated decision-making in collaborative settings.

Figures

Figures reproduced from arXiv: 2506.20373 by the authors.

Figure 1
Figure 1. Overview of the proposed framework. The Action Reasoner is a VLM (GPT4o [19]) whose prompt has three inputs: (1) A region of the current [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Visualization of the prompt provided to VLM. The prompt begins with a general introduction, followed by cropped object images, each uniquely [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 14 canonical work pages

  1. [1]

    Large language models for human–robot interaction: A review,

    C. Zhang, J. Chen, J. Li, Y . Peng, and Z. Mao, “Large language models for human–robot interaction: A review,” Biomimetic Intel- ligence and Robotics, vol. 3, no. 4, p. 100131, Dec. 2023, doi: 10.1016/j.birob.2023.100131

    arXiv 2023

  2. [2]

    Benchmark Eval- uations, Applications, and Challenges of Large Vision Language Models: A Survey,

    Z. Li, X. Wu, H. Du, H. Nghiem, and G. Shi, “Benchmark Eval- uations, Applications, and Challenges of Large Vision Language Models: A Survey,” Jan. 10, 2025, arXiv: arXiv:2501.02189. doi: 10.48550/arXiv.2501.02189

    arXiv doi:10.48550/arxiv.2501.02189 2025

  3. [3]

    MUTEX: Learning Unified Policies from Multimodal Task Specifications,

    R. Shah, R. Mart ´ın-Mart´ın, and Y . Zhu, “MUTEX: Learning Unified Policies from Multimodal Task Specifications,” Sep. 25, 2023, arXiv: arXiv:2309.14320. doi: 10.48550/arXiv.2309.14320

    arXiv doi:10.48550/arxiv.2309.14320 2023

  4. [4]

    Vision- language model-driven scene understanding and robotic object manip- ulation,

    S. Liu, J. Zhang, R. X. Gao, X. Vincent Wang, and L. Wang, “Vision- language model-driven scene understanding and robotic object manip- ulation,” in 2024 IEEE 20th International Conference on Automation Science and Engineering (CASE), Bari, Italy: IEEE, Aug. 2024, pp. 21–26. doi: 10.1109/CASE59546.2024.10711845

    arXiv 2024

  5. [6]

    CoPAL: Corrective Planning of Robot Actions with Large Language Models

    Joublin, Frank, et al. CoPAL: Corrective Planning of Robot Actions with Large Language Models. arXiv:2310.07263, arXiv, 11 Oct. 2023, https://doi.org/10.48550/arXiv.2310.07263

    work page Pith review arXiv doi:10.48550/arxiv.2310.07263 2023

  6. [7]

    VLM-Social-Nav: Socially Aware Robot Navigation through Scoring using Vision-Language Models,

    D. Song, J. Liang, A. Payandeh, A. H. Raj, X. Xiao, and D. Manocha, “VLM-Social-Nav: Socially Aware Robot Navigation through Scoring using Vision-Language Models,” Nov. 25, 2024, arXiv: arXiv:2404.00210. doi: 10.48550/arXiv.2404.00210

    arXiv doi:10.48550/arxiv.2404.00210 2024

  7. [8]

    VLFM: Vision-Language Frontier Maps for Zero-Shot Seman- tic Navigation,

    N. Yokoyama, S. Ha, D. Batra, J. Wang, and B. Bucher, “VLFM: Vision-Language Frontier Maps for Zero-Shot Seman- tic Navigation,” Dec. 06, 2023, arXiv: arXiv:2312.03275. doi: 10.48550/arXiv.2312.03275

    arXiv doi:10.48550/arxiv.2312.03275 2023

  8. [9]

    ZSON: Zero-Shot Object-Goal Navigation using Multimodal Goal Embeddings,

    A. Majumdar, G. Aggarwal, B. Devnani, J. Hoffman, and D. Ba- tra, “ZSON: Zero-Shot Object-Goal Navigation using Multimodal Goal Embeddings,” Oct. 13, 2023, arXiv: arXiv:2206.12403. doi: 10.48550/arXiv.2206.12403

    arXiv doi:10.48550/arxiv.2206.12403 2023

Show all 29 references
  1. [10]

    LaMI: Large Language Models for Multi-Modal Human-Robot Interaction,

    C. Wang et al., “LaMI: Large Language Models for Multi-Modal Human-Robot Interaction,” in Extended Abstracts of the CHI Con- ference on Human Factors in Computing Systems, May 2024, pp. 1–10. doi: 10.1145/3613905.3651029

    2024

  2. [11]

    To Help or Not to Help: LLM-based Attentive Support for Human-Robot Group Interactions,

    D. Tanneberg et al., “To Help or Not to Help: LLM-based Attentive Support for Human-Robot Group Interactions,” in 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Abu Dhabi, United Arab Emirates: IEEE, Oct. 2024, pp. 9130–9137. doi: 10.1109/IROS...

    2024

  3. [12]

    VLM See, Robot Do: Human Demo Video to Robot Action Plan via Vision Language Model,

    B. Wang, J. Zhang, S. Dong, I. Fang, and C. Feng, “VLM See, Robot Do: Human Demo Video to Robot Action Plan via Vision Language Model,” Oct. 11, 2024, arXiv: arXiv:2410.08792. doi: 10.48550/arXiv.2410.08792

    2024 doi

  4. [13]

    Robots Can Multitask Too: Integrating a Mem- ory Architecture and LLMs for Enhanced Cross-Task Robot Ac- tion Generation,

    H. Ali et al., “Robots Can Multitask Too: Integrating a Mem- ory Architecture and LLMs for Enhanced Cross-Task Robot Ac- tion Generation,” Oct. 11, 2024, arXiv: arXiv:2407.13505. doi: 10.48550/arXiv.2407.13505

    2024 doi arXiv

  5. [14]

    “Exploring large language models as a source of common-sense knowledge for robots“

    Felix Ocker and Joerg Deigmoeller and Julian Eggert. “Exploring large language models as a source of common-sense knowledge for robots“. International Semantic Web Conference (ISWC), 2023

    2023

  6. [15]

    Takuma Yoneda and Jiading Fang and Peng Li and Huanyu Zhang and Tianchong Jiang and Shengjie Lin and Ben Picker and David Yunis and Hongyuan Mei and Matthew R., “Statler: State-Maintaining Language Models for Embodied Reasoning“, Proceedings of the IEEE International Conferenc...

    2024

  7. [16]

    F. Liu, K. Fang, P. Abbeel, S. Levine, “MOKA: Open-World Robotic Manipulation through Mark-Based Visual Prompting“, Robotics: Sci- ence and Systems (RSS), 2024

    2024

  8. [17]

    “Quo vadis, action recogni- tion? A new model and the kinetics dataset.“ Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017

    Carreira, Joao, and Andrew Zisserman. “Quo vadis, action recogni- tion? A new model and the kinetics dataset.“ Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017

    2017

  9. [19]

    Achiam et al

    J. Achiam et al. “GPT-4 Technical Report“. 2024, arXiv:2405.21075. doi: 10.48550/arXiv.2405.21075.2024, arXiv:2303.08774. doi:10.48550/arXiv.2303.08774

    2024 doi arXiv

  10. [20]

    Agrawal et al

    P. Agrawal et al. “Pixtral 12B“. 2024, arXiv:2410.07073. doi: doi:10.48550/arXiv.2303.08774

    2024 doi arXiv

  11. [21]

    and Li, D

    Li, J. and Li, D. and Savarese, S. and Hoi, S.. “BLIP-2: Boot- strapping Language-Image Pre-training with Frozen Image En- coders and Large Language Models“. 2023, arXiv:2301.12597. doi: 10.48550/arXiv.2301.12597

    2023 doi arXiv

  12. [22]

    “LLaMA: Open and Efficient Foundation Language Models“ arXiv:2302.13971, 2023

    Touvron, H et al. “LLaMA: Open and Efficient Foundation Language Models“ arXiv:2302.13971, 2023

    2023 arXiv

  13. [23]

    and Kembhavi, A., 2020

    Pratt, S., Yatskar, M., Weihs, L., Farhadi, A. and Kembhavi, A., 2020. “Grounded situation recognition“. European Conference on Computer Vision (ECCV), 2020

    2020

  14. [24]

    and Chen, L

    Lei, J., Li, L., Wang, C., Xiao, J. and Chen, L. “Seeing beyond classes: Zero-shot grounded situation recognition via language explainer“. In Proceedings of the 32nd ACM International Conference on Multime- dia, 2024

    2024

  15. [25]

    and Saffiotti, A

    Coradeschi, s. and Saffiotti, A. “An introduction to the anchoring problem“. Robotics and autonomous systems, 2003

    2003

  16. [26]

    and Ros, R

    Lemaignan, S. and Ros, R. and Alami, R. and Beetz, M. “What are you talking about? Grounding dialogue in a perspective-aware robotic architecture“ Conference on Robot and Human Interactive Communication (RO-MAN), 2011

    2011

  17. [27]

    and Deigmoeller, J

    Eggert, J. and Deigmoeller, J. and Fischer L. and Richter, A. “Action representation for intelligent agents using memory nets“. International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management, 2019

    2019

  18. [28]

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context,

    M. Reid, N. Savinov, D. Teplyashin, D. Lepikhin, T. Lillicrap, J.-b. Alayrac, R. Soricut, A. Lazaridou, O. Firat, J. Schrittwieser, et al., “Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context,” arXiv:2403.05530, 2024

    2024 arXiv

  19. [29]

    Llava-onevision: Easy visual task transfer,

    B. Li, Y . Zhang, D. Guo, R. Zhang, F. Li, H. Zhang, K. Zhang, Y . Li, Z. Liu, and C. Li, “Llava-onevision: Easy visual task transfer,” arXiv:2408.03326, 2024

    2024 arXiv

  20. [30]

    Vila: On pre-training for visual language models,

    J. Lin, H. Yin, W. Ping, P. Molchanov, M. Shoeybi, and S. Han, “Vila: On pre-training for visual language models,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024

    2024

  21. [31]

    Kr `‘uger, C

    N. Kr `‘uger, C. Geib, J. Piater, R. Petrik, M. Steedman, F. W `‘org`‘otter, A. Ude, T. Asfour, D. Kraft, D. Omr ˇcen, A. Agostinig, R. Dillmann, “Object–action complexes: Grounded abstractions of sensory–motor processes“. Robotics and Autonomous Systems, volume 59, number 10,...

    2011

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.