Pith. sign in

REVIEW 4 major objections 7 minor 38 references

Goal-oriented Communication for Fast and Robust Robotic Fault Detection and Recovery

T0 review · 4 major / 7 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Transmitting compact 3D scene graphs instead of raw images cuts robotic fault-detection and recovery time by up to 82.6% and lifts task success by up to 76% in simulations.

desk verdict Plausible system-level integration of 3D scene graphs, edge-point digital twins, and a distilled SLM for robotic FDR, but the headline numbers rest on an unreleased simulator and an unvalidated relation classifier. read the letter →

arxiv 2601.18765 v2 pith:F57WGM24 submitted 2026-01-26 cs.RO

Shutong Chen , Adnan Aijaz , Yansha Deng This is my paper
classification cs.RO
keywords goal-orientedcommunicationroboticfaultdetectionandrecovery3Dscenegraphsemanticrepresentationsmalllanguagemodeldigitaltwinreconstructionedgepointsamplingwirelessoffloading
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

This paper tries to establish that robotic fault detection and recovery should be driven end-to-end by the downstream goal: send only semantic representations—a 3D scene graph for task-level faults, object edge points for motion-level faults—and replace general-purpose large models with a fine-tuned small language model for recovery. It argues and simulates that this jointly designed communication-computation-control loop reduces fault detection and recovery time by up to 82.6% and improves task success rate by up to 76% relative to vision-language and large-language baselines across three industrial manipulation tasks. A sympathetic reader would care because it suggests that fault recovery in wireless robot systems can be both faster than human reaction times and more reliable by treating communication as task-goal-driven rather than raw-data transport.

What carries the argument

The dual-branch semantic representation extractor. For fault detection it builds a 3D scene graph—an undirected graph whose nodes are objects and whose edges are predicted spatial relations such as 'standing on' or 'next to'—using object segmentation, point-cloud encoding, and a graph convolutional network trained with cross-entropy on an indoor scene-graph dataset. For motion-level faults it samples attention-based edge points and fits smooth curves to reconstruct lightweight object contours, enabling digital-twin verification. Fault detection is thereby reduced to checking whether expected relation transitions appear or unexpected proximity relations emerge, and transmission happens only w

What would settle it

Run the fault-detection stage in a cluttered scene with partial occlusions and measure precision and recall for the spatial-relation predictions against ground truth. A concrete test: compare task success when recovery uses the learned 3D scene graph versus a perfect, oracle-provided graph; if success collapses without the oracle, the bottleneck is detection, not communication.

Watch

Extended reading notes

Core claim

The central claim is that a goal-oriented communication loop, where an RGB-D camera locally builds a 3D scene graph of objects and their spatial relationships, transmits it only when a fault is detected, and task-level recovery is generated by a distilled small language model—with optional digital-twin refinement from transmitted object edge points for motion-level faults—dramatically outperforms the state of the art: up to 82.6% reduction in fault detection and recovery time and up to 76% higher task success rate across three simulated industrial tasks. The paper also claims that adaptive bandwidth thresholds optimally choose local versus edge computation, and that knowledge distillation le

Load-bearing premise

The framework assumes that a graph-convolution classifier, trained on indoor scenes but deployed in simulated robot workspaces with hand-set distance thresholds, correctly predicts the spatial relations that define faults; if relation prediction is inaccurate, fault detection fails regardless of communication or model speedups.

Editorial extensions

If this is right

  • Task-level faults can be detected without any vision-language inference per frame, replacing periodic image transmission with a compact 3D scene graph and shrinking detection time substantially.
  • A fine-tuned small language model enlivened by knowledge distillation can replace large language models for recovery-motion generation, matching success rates while cutting inference time by roughly 92% according to the ablation.
  • Object edge points plus lightweight digital twin reconstruction improve motion-level task success by about 28% by simulating and refining recovery trajectories before execution.
  • Two bandwidth thresholds give a practical decision rule: below the lower threshold compute locally and send only semantics; above the upper threshold send full point clouds; in between, use a hybrid strategy.
  • Combining these pieces can push overall fault detection and recovery below the roughly 500 ms human-reaction window considered relevant for human-robot collaboration.

Reading between the lines

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

  • The headline gains rest on simulations; the largest open question is detection accuracy of the spatial-relation classifier in cluttered real scenes, since every downstream benefit depends on the 3D scene graph being correct.
  • The same 'transmit only task-relevant semantic structure' principle could extend to other robot tasks—navigation, assembly, or human-robot handover—where the bottleneck is wireless uplink of raw sensor data rather than the reasoning itself.
  • Knowledge distillation could be made an ongoing online loop, letting the small model improve on-site from a large teacher or from real execution feedback instead of a fixed offline dataset.
  • A testable extension is to replace the hand-set distance threshold used to prune scene-graph edges with a learned or bandwidth-aware threshold, and measure whether both detection latency and task success improve.
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

4 major / 7 minor

Summary. The paper proposes a Goal-oriented Communication (GoC) framework for robotic fault detection and recovery (FDR) that transmits a 3D scene graph (3D-SG) for fault detection and object edge points for lightweight digital twin reconstruction, and uses a fine-tuned small language model (SLM) with knowledge distillation for recovery. The framework is evaluated in MuJoCo simulations across three industrial tasks, reporting up to 82.6% reduction in FDR time and up to 76% higher task success rate compared to VLM/LLM-based SOTA baselines.

Significance. If the results hold, the paper makes a useful contribution by demonstrating that task-oriented semantic representations can substantially reduce communication and inference latency in robotic FDR, and that a fine-tuned SLM can approach LLM-level recovery performance. The combination of 3D-SG, edge-point digital twin, and knowledge distillation is novel, and the ablation study and adaptive offloading strategy add value. However, the lack of detection validation in the deployment setting, missing absolute timing breakdowns, absent statistical variance, and the self-referential training/evaluation limit confidence in the reported improvements.

major comments (4)
  1. [§IV-A, §III-C, §V] The fault detector is not validated in the deployment setting. The TripletGCN is trained on 3RScan (Eq. 10) but applied to MuJoCo scenes with industrial objects (workpieces, parcels, pallets) not present in 3RScan. Section III-C defines fault detection via hand-defined expected relation transitions and a fixed distance threshold. No detection accuracy, precision/recall, confusion matrix, or sensitivity analysis is reported for the actual MuJoCo scenes. The headline reductions (82.6% FDR time, 76% success) assume the relation classifier is reliable; false positives would inflate FDR time and false negatives would degrade success rate. Please add detection-evaluation metrics in the simulation environment or a sensitivity analysis over the threshold.
  2. [§V, Eqs. (5)-(6)] The paper reports only percentage reductions in detection/recovery time, with no absolute timing values or per-component breakdown. The claimed 82.6% reduction cannot be assessed without knowing the contributions of t_sg, t_com, t_inf, and t_exe. Provide a breakdown table (e.g., per task and framework) with raw time values so the reader can verify the communication and computation savings.
  3. [§V] Success-rate differences are reported without variance or significance testing. Only 25 runs per task are mentioned; no confidence intervals, standard deviations, or tests are provided. Since the differences (e.g., 12%–76%) are large, they may be significant, but the paper should show this. Also, no code or data are released (only a demo link), which hampers reproducibility.
  4. [§IV-D] The training and evaluation are self-referential. The fine-tuning dataset in Sec. IV-D.1 is generated by GPT-4o, which is also the teacher in the knowledge-distillation module (Sec. IV-D.2). The evaluation uses the same MuJoCo tasks and fault injections that likely generated this data. This risks overfitting to the teacher and to the simulator. Please specify how the evaluation fault scenarios differ from the training distribution, or evaluate on held-out tasks/faults.
minor comments (7)
  1. [§IV-A, Eqs. (12)-(13)] The same symbol α is used for the raw and normalized attention scores, causing confusion. Rename one of them.
  2. [§IV-D, Eq. (21)] The LoRA scaling is written as r/α·ΔW, while standard LoRA uses α/r·ΔW. Please check the notation and the surrounding text near Eq. (21).
  3. [§V-E heading] Typo: 'Adaptative' should be 'Adaptive'.
  4. [§III-D, Eq. (3)] The channel model uses Nakagami-m fading with power gain g, but Eq. (3) takes E[|g|^2], which is inconsistent with the definition. Clarify.
  5. [§I contribution bullet] Typo: 'Mujoco' should be 'MuJoCo'.
  6. [Fig. 3] The prompt example is referenced in §IV-D.1 but not visible in the text; ensure the figure is included.
  7. [§III-C] The claim that digital twin reconstruction and verification time is negligible (<2 ms) is unsubstantiated. Add a measurement or a justification.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the headline FDR-time and success-rate figures are measured in simulation against external baselines, and no claimed result reduces to a fitted parameter or a self-citation chain.

full rationale

The paper's central quantitative claims (82.6% FDR-time reduction, up to 76% success-rate improvement) are empirical outcomes of MuJoCo experiments (Sec. V), not consequences of the model equations by construction. The TripletGCN is trained on external 3RScan labels via cross-entropy (Eq. 10) and then applied to the simulated scenes; the test-time task success is determined by the simulator, not by the training loss. The adaptive-offloading thresholds (Sec. IV-C) are defined by equating t_det_local and t_det_edge; Sec. V-E's offloading conclusions are therefore algebraic consequences of the decision rule, not independent predictions, but this is a formal identity in an auxiliary module rather than a circular derivation of the headline results. The self-citations (refs [7], [11], [20]) motivate the goal-oriented-communication and digital-twin ideas but are not load-bearing for the simulation results. Sec. II.A itself notes the difficulty of extracting 3D-SG from noisy point clouds, and the fine-tuning/KD procedure (Sec. IV-D) uses GPT-4o both as data generator and online teacher with simulator-based validation (footnote 2); these raise external-validity and evaluation-confound concerns, but they are not instances of a claimed prediction being equivalent to its inputs by definition or equation. No circular step satisfying the evidence threshold was found.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

No new physical entities are introduced. The framework relies on domain assumptions about simulation-to-real transfer, scene-graph detection reliability, GPT-4o-generated training ground truth, and baseline fairness. Free parameters are largely unspecified (distance threshold, λ, LoRA rank), which weakens reproducibility but does not constitute invented entities.

free parameters (5)
  • 3D-SG edge distance threshold = not specified
    Edges whose spatial distance exceeds a predefined threshold are removed (Sec. IV-A); value not given, directly controls graph structure and hence fault detection.
  • TopK number of edge points (512) = 512
    Number of edge points selected per object (Sec. IV-B1); chosen by hand, affects communication size and digital twin fidelity.
  • Curve-fitting regularization coefficient λ = not specified
    Controls B-spline smoothness vs. deviation (Eq. 16); tuned without stated method or value.
  • LoRA rank r and scaling α = not fully specified (rank implied small)
    LoRA hyperparameters (Sec. IV-D1) control fine-tuning capacity; table gives scaling factor 16 but not rank.
  • Distance threshold for keypoint redundancy in baseline = not specified
    Affects the keypoint baseline, not the central claim directly, but used in the comparison.
assumptions (6)
  • domain assumption MuJoCo simulation accurately represents real industrial FDR dynamics
    All FDR time and success-rate claims are computed in MuJoCo (Sec. V-A); transfer to physical robots is unvalidated.
  • domain assumption TripletGCN trained on 3RScan generalizes to the simulated workspaces without fine-tuning or accuracy evaluation
    Sec. IV-A2 trains the relation classifier on 3RScan but never reports detection accuracy in the MuJoCo scenes, despite detection being the trigger for all downstream recovery.
  • domain assumption Spatial relation changes in 3D-SG are sufficient to distinguish task-level vs motion-level faults
    Sec. III-C defines fault types by expected vs observed scene-graph transitions; this assumes a fixed, hand-defined mapping of relations to faults that is complete for the task.
  • domain assumption GPT-4o-generated recovery motions with manual refinement form ground truth for SLM fine-tuning
    Sec. IV-D1 constructs the dataset from 'running robotic tasks simulations online and prompting GPT-4o'; the teacher model defines correctness, so SLM quality is bounded by GPT-4o and the manual refinement.
  • domain assumption Zero-shot/one-shot constraint generation by a 90B VLM/LLM is a fair SOTA baseline for FDR
    Baselines [4],[6] are adapted without their original fine-tuning; the comparison may understate baseline success rates. The paper does not report fine-tuned baseline variants.
  • domain assumption Digital twin verification time is negligible
    Sec. III-E states twin reconstruction/verification 'typically incur negligible computing latency (typically under 2ms)' but later ablation (Fig. 6b) attributes a 75% SLM inference increase to verification rounds, partially contradicting the 'negligible' assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Goal-oriented Communication for Fast and Robust Robotic Fault Detection and Recovery." pith.science (2026). https://pith.science/paper/F57WGM24

@misc{pith2026260118765,
  author       = {Pith},
  title        = {Pith review of: Goal-oriented Communication for Fast and Robust Robotic Fault Detection and Recovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F57WGM24}},
  note         = {Machine review of arXiv:2601.18765}
}
read the original abstract

Autonomous robotic systems are widely deployed in smart factories and operate in dynamic, uncertain, and human-involved environments that require low-latency and robust fault detection and recovery (FDR). However, existing FDR frameworks exhibit various limitations, such as significant delays in communication and computation, and unreliability in robot motion/trajectory generation, mainly because the communication-computation-control (3C) loop is designed without considering the downstream FDR goal. To address this, we propose a novel Goal-oriented Communication (GoC) framework that jointly designs the 3C loop tailored for fast and robust robotic FDR, with the goal of minimising the FDR time while maximising the robotic task (e.g., workpiece sorting) success rate. For fault detection, our GoC framework innovatively defines and extracts the 3D scene graph (3D-SG) as the semantic representation via our designed representation extractor, and detects faults by monitoring spatial relationship changes in the 3D-SG. For fault recovery, we fine-tune a small language model (SLM) via Low-Rank Adaptation (LoRA) and enhance its reasoning and generalization capabilities via knowledge distillation to generate recovery motions for robots. We also design a lightweight goal-oriented digital twin reconstruction module to refine the recovery motions generated by the SLM when fine-grained robotic control is required, using only task-relevant object contours for digital twin reconstruction. Extensive simulations demonstrate that our GoC framework reduces the FDR time by up to 82.6% and improves the task success rate by up to 76%, compared to the state-of-the-art frameworks that rely on vision language models for fault detection and large language models for fault recovery.

Figures

Figures reproduced from arXiv: 2601.18765 by the authors.

Figure 1
Figure 1. State-of-the-art Frameworks and Our GoC Framework for Robotic Fault Detection and Recovery. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The FDR time comparison. A. Scene Graph Generation Robotic fault detection fundamentally relies on understand￾ing the spatial relationships among objects in the workspace. To this end, we propose to use the more direct and efficient 3D-SG to represent the workspace spatial relationships and enable lightweight wireless transmission. For all the objects Oi = {O1 i , . . . , ONO i i } observed in image frame i, we firs… view at source ↗
Figure 3
Figure 3. Fine-tuning prompt example. However, for motion-level faults that require more bandwidth￾intensive edge points, the edge server’s superior computing power starts to compensate for transmission delay, which means it is preferable to adopt a hybrid offloading strategy that performs local computing for task-level faults and delegates motion-level fault processing to edge server. When bandwidth exceeds B2, the enhanced … view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparison of FDR time and task success rate between SOTA and the proposed GoC frameworks in handling task-level faults. [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Comparison of FDR time and task success rate between SOTA and the proposed GoC frameworks in handling motion-level faults. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Ablation Study of our proposed GoC framework. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Effectiveness of our proposed adaptative offloading scheme. [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 6 linked inside Pith

  1. [1]

    Modeling human response to robot errors for timely error detection,

    M. Stiber, R. Taylor, and C.-M. Huang, “Modeling human response to robot errors for timely error detection,” inProc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), Kyoto, Japan, Dec. 2022, pp. 676–683

    2022

  2. [2]

    Online replanning in belief space for partially observable task and motion problems,

    C. R. Garrettet al., “Online replanning in belief space for partially observable task and motion problems,” inProc. IEEE Int. Conf. Robot. Autom. (ICRA), Paris, France, May 2020, pp. 5678–5684

    2020

  3. [3]

    Modular multi-level replanning tamp framework for dynamic environment,

    T. Lin, C. Yue, Z. Liu, and X. Cao, “Modular multi-level replanning tamp framework for dynamic environment,”IEEE Robot. Autom. Lett., vol. 9, no. 5, pp. 4234–4241, Mar. 2024

    2024

  4. [4]

    DoReMi: Grounding language model by detecting and recovering from plan-execution misalignment,

    Y . Guoet al., “DoReMi: Grounding language model by detecting and recovering from plan-execution misalignment,” inProc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), Abu Dhabi, United Arab Emirates, Dec. 2024, pp. 12 124–12 131

    2024

  5. [5]

    Copa: General robotic manipulation through spatial constraints of parts with foundation models,

    H. Huanget al., “Copa: General robotic manipulation through spatial constraints of parts with foundation models,” inProc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), Dec. 2024, pp. 9488–9495

    2024

  6. [6]

    Rekep: Spatio-temporal reasoning of relational key- point constraints for robotic manipulation,

    W. Huanget al., “Rekep: Spatio-temporal reasoning of relational key- point constraints for robotic manipulation,” inProc. Conf. Robot Learn. (CoRL), Munich, Germany, Nov. 2024

    2024

  7. [7]

    Goal- oriented semantic communications for 6G networks,

    H. Zhou, Y . Deng, X. Liu, N. Pappas, and A. Nallanathan, “Goal- oriented semantic communications for 6G networks,”IEEE Internet Things Mag., vol. 7, no. 5, pp. 104–110, Sept. 2024

    2024

  8. [8]

    Mujoco: A physics engine for model- based control,

    E. Todorov, T. Erez, and Y . Tassa, “Mujoco: A physics engine for model- based control,” inProc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), Vilamoura-Algarve, Portugal, Dec. 2012, pp. 5026–5033

    2012

Show all 38 references
  1. [9]

    Beyond transmitting bits: Context, semantics, and task-oriented communications,

    D. G ¨und¨uzet al., “Beyond transmitting bits: Context, semantics, and task-oriented communications,”IEEE J. Sel. Areas Commun., vol. 41, no. 1, pp. 5–41, Jan. 2023

    2023

  2. [10]

    Goal-oriented semantic communication for wireless visual question answering,

    S. Liuet al., “Goal-oriented semantic communication for wireless visual question answering,”arXiv preprint arXiv:2411.02452, Nov. 2024

    2024 arXiv

  3. [11]

    Goal-oriented semantic communications for avatar-centric augmented reality,

    Z. Wang, Y . Deng, and A. Hamid Aghvami, “Goal-oriented semantic communications for avatar-centric augmented reality,”IEEE Trans. Commun., vol. 72, no. 12, pp. 7982–7995, Jul. 2024

    2024

  4. [12]

    Task-oriented image transmission for scene classification in unmanned aerial systems,

    X. Kang, B. Song, J. Guo, Z. Qin, and F. R. Yu, “Task-oriented image transmission for scene classification in unmanned aerial systems,”IEEE Trans. Commun., vol. 70, no. 8, pp. 5181–5192, Jun. 2022

    2022

  5. [13]

    Toward semantic communications: Deep learning-based image semantic coding,

    D. Huang, F. Gao, X. Tao, Q. Du, and J. Lu, “Toward semantic communications: Deep learning-based image semantic coding,”IEEE J. Sel. Areas Commun., vol. 41, no. 1, pp. 55–71, Nov. 2023

    2023

  6. [14]

    Roco: Dialectic multi-robot collab- oration with large language models,

    Z. Mandi, S. Jain, and S. Song, “Roco: Dialectic multi-robot collab- oration with large language models,” inProc. IEEE Int. Conf. Robot. Autom. (ICRA), Yokohama, Japan, May. 2024, pp. 286–299

    2024

  7. [15]

    Smart-llm: Smart multi-agent robot task planning using large language models,

    S. S. Kannan, V . L. N. Venkatesh, and B.-C. Min, “Smart-llm: Smart multi-agent robot task planning using large language models,” inProc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), Abu Dhabi, United Arab Emirates, Dec. 2024, pp. 12 140–12 147

    2024

  8. [16]

    Llm-planner: Few-shot grounded planning for embodied agents with large language models,

    C. H. Songet al., “Llm-planner: Few-shot grounded planning for embodied agents with large language models,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), Paris, France, Oct. 2023, pp. 2998–3009

    2023

  9. [17]

    Fastnav: Fine-tuned adaptive small- language- models trained for multi-point robot navigation,

    Y . Chen, Y . Han, and X. Li, “Fastnav: Fine-tuned adaptive small- language- models trained for multi-point robot navigation,”IEEE Robot. Autom. Lett., vol. 10, no. 1, pp. 390–397, Nov. 2024

    2024

  10. [18]

    Can only llms do reasoning?: Potential of small language models in task planning,

    G. Choiet al., “Can only llms do reasoning?: Potential of small language models in task planning,”arXiv preprint arXiv:2404.03891, Apr. 2024

    2024 arXiv

  11. [19]

    Digital twin in industry: State-of-the-art,

    F. Taoet al., “Digital twin in industry: State-of-the-art,”IEEE Trans. Ind. Informat., vol. 15, no. 4, pp. 2405–2415, Apr. 2019

    2019

  12. [20]

    Goal-oriented semantic communication for robot arm reconstruction in digital twin: Feature and temporal selections,

    S. Chenet al., “Goal-oriented semantic communication for robot arm reconstruction in digital twin: Feature and temporal selections,”IEEE J. Sel. Areas Commun., pp. 1–1, May 2025

    2025

  13. [21]

    Multi-UA V cooperative search based on reinforcement learning with a digital twin driven training framework,

    G. Shenet al., “Multi-UA V cooperative search based on reinforcement learning with a digital twin driven training framework,”IEEE Trans. Veh. Technol., vol. 72, no. 7, pp. 8354–8368, Feb. 2023

    2023

  14. [22]

    Digital twin-based autonomous navigation and control of omnidirectional mobile robots,

    H. Yang, Z. Qin, Y . Xia, and F. Cheng, “Digital twin-based autonomous navigation and control of omnidirectional mobile robots,”IEEE Trans. Veh. Technol., vol. 74, no. 4, pp. 5687–5697, Apr. 2025

    2025

  15. [23]

    Toward communication-efficient digital twin via AI- powered transmission and reconstruction,

    M. Liet al., “Toward communication-efficient digital twin via AI- powered transmission and reconstruction,”IEEE J. Sel. Areas Commun., vol. 41, no. 11, pp. 3624–3635, Nov. 2023

    2023

  16. [24]

    Sampling, communication, and prediction co-design for synchronizing the real-world device and digital model in metaverse,

    Z. Menget al., “Sampling, communication, and prediction co-design for synchronizing the real-world device and digital model in metaverse,” IEEE J. Sel. Areas Commun., vol. 41, no. 1, pp. 288–300, Jan. 2023

    2023

  17. [25]

    Automated creation of digital cousins for robust policy learning,

    T. Daiet al., “Automated creation of digital cousins for robust policy learning,” inProc. Conf. Robot Learn. (CoRL), Munich, Germany, Nov. 2024

    2024

  18. [26]

    Evaluation of the azure kinect and its comparison to kinect v1 and kinect v2,

    M. T ¨olgyessyet al., “Evaluation of the azure kinect and its comparison to kinect v1 and kinect v2,”Sensors, vol. 21, no. 2, p. 413, Jan. 2021

    2021

  19. [27]

    Learn-gen-plan: Bridging the gap between vision language models and real-world long-horizon dexterous manipulations,

    P. Haoet al., “Learn-gen-plan: Bridging the gap between vision language models and real-world long-horizon dexterous manipulations,”IEEE Trans. Autom. Sci. Eng., vol. 22, pp. 15 638–15 649, May 2025

    2025

  20. [28]

    DINOv2: Learning robust visual features without supervision,

    M. Oquabet al., “DINOv2: Learning robust visual features without supervision,”arXiv preprint arXiv:2304.07193, Feb. 2024

    2024 arXiv

  21. [29]

    Segment anything,

    A. Kirillovet al., “Segment anything,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), Paris, France, Oct. 2023, pp. 4015–4026

    2023

  22. [30]

    YOLOv11: An overview of the key archi- tectural enhancements,

    R. Khanam and M. Hussain, “YOLOv11: An overview of the key archi- tectural enhancements,”arXiv preprint arXiv:2410.17725, Oct. 2024

    2024 arXiv

  23. [31]

    Technical Specification Group Radio Access Network; Study on channel model for frequencies from 0.5 to 100 GHz (Release 16),

    3GPP, “Technical Specification Group Radio Access Network; Study on channel model for frequencies from 0.5 to 100 GHz (Release 16),” TR 38.901 V16.1.0, Dec. 2019

    2019

  24. [32]

    Incremental 3d semantic scene graph prediction from rgb sequences,

    S.-C. Wuet al., “Incremental 3d semantic scene graph prediction from rgb sequences,” inProc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit. (CVPR), Vancouver, BC, Canada, Jun. 2023, pp. 5064–5074

    2023

  25. [33]

    PointNeXt: Revisiting PointNet++ with improved training and scaling strategies,

    G. Qianet al., “PointNeXt: Revisiting PointNet++ with improved training and scaling strategies,” inProc. Int. Conf. Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 35, New Orleans, USA, Nov. 2022, pp. 23 192–23 204

    2022

  26. [34]

    RIO: 3D object instance re-localization in changing indoor environments,

    J. Waldet al., “RIO: 3D object instance re-localization in changing indoor environments,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), Seoul, Korea, Oct. 2019, pp. 7657–7666

    2019

  27. [35]

    Attention-based point cloud edge sampling,

    C. Wuet al., “Attention-based point cloud edge sampling,” inProc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit. (CVPR), Vancouver, BC, Canada, Jun. 2023, pp. 5333–5343

    2023

  28. [36]

    The Llama 3 herd of models,

    A. Grattafioriet al., “The Llama 3 herd of models,”arXiv preprint arXiv:2407.21783, Nov. 2024

    2024 arXiv

  29. [37]

    LoRA: Low-rank adaptation of large language models,

    E. J. Huet al., “LoRA: Low-rank adaptation of large language models,” arXiv preprint arXiv:2106.09685, Oct. 2022

    2022 arXiv

  30. [38]

    Dense robotic packing of irregular and novel 3d objects,

    F. Wanget al., “Dense robotic packing of irregular and novel 3d objects,” IEEE Trans. Robot., vol. 38, no. 2, pp. 1160–1173, Aug. 2022

    2022

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.