Pith. sign in

REVIEW 4 major objections 6 minor 39 references

A wearable suite of two Apple Watches and capacitive gloves recognizes 20 drone-control gestures as accurately as a camera-based method, while using far less computation and model size.

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 20:10 UTC pith:JBXASCEZ

load-bearing objection Useful dataset and honest ablations, but the sensor-vs-vision parity claim needs equivalence testing and more than six participants. the 4 major comments →

arxiv 2602.23694 v3 pith:JBXASCEZ submitted 2026-02-27 cs.RO cs.AI

Interpretable Multimodal Gesture Recognition for Drone and Mobile Robot Teleoperation via Log-Likelihood Ratio Fusion

Seungyeol Baek , Jaspreet Singh , Lala Shakti Swarup Ray , Hymalai Bello , Paul Lukowicz , Sungho Suh This is my paper
classification cs.RO cs.AI
keywords multimodal gesture recognitionlog-likelihood ratio fusionwearable sensorsdrone and robot teleoperationaircraft marshalling gesturesinertial measurement unitscapacitive sensingmodel interpretability
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.

This paper claims that hands-free teleoperation of drones and mobile robots can be made as accurate as camera-based recognition without using cameras at all. The authors build a system from Apple Watch inertial sensors on both wrists plus custom capacitive gloves, and fuse the eight sensor streams with a log-likelihood ratio (LLR) late-fusion layer. On a new dataset of 20 aircraft-marshalling-inspired gestures, the fused sensor model reaches 95.40% macro F1 in a leave-one-session-out split and 93.59% in a leave-one-participant-out split, matching or slightly exceeding the vision baseline PoseConv3D while needing a fraction of its compute, model size, and training time. The LLR fusion also reports each sensor's per-class contribution, so an operator can see why a command was issued. A sympathetic reader would take the central claim to be that wearable-only, interpretable sensing is a viable replacement for vision in gesture-driven robot control, especially in environments where cameras fail.

Core claim

The central discovery is that late fusion by log-likelihood ratios lets heterogeneous wearable sensors—wrist accelerometers, gyroscopes, quaternions, and glove capacitance—recognize a 20-gesture teleoperation vocabulary with accuracy comparable to a state-of-the-art vision-based pose-recognition model. In the LOSO split the LLR-fused sensor model scores F1 95.40±5.91 versus PoseConv3D's 94.70±2.33; in the harder LOPO split it scores 93.59±5.20 versus 93.39±5.44. Each of the eight modality streams (left/right wrist × four sensor types) is encoded by a CNN+GRU, and the per-modality softmax probabilities are converted to log-likelihood ratios and summed before a final linear classifier. This co

What carries the argument

The load-bearing mechanism is the log-likelihood ratio (LLR) fusion layer. For each of the eight sensor-modality streams, a CNN+GRU encoder produces a feature vector; a softmax over the 21 classes yields a probability per class; the LLR for class i is ln(p_i / (1 − p_i)), the log odds of that modality's support for the class. Summing these log-odds across all eight streams produces the fused representation, which is passed to a single fully connected layer for classification. The sum is linear, so each modality's contribution remains explicit and signed, which is what makes the fusion interpretable. The paper also evaluates a self-attention fusion variant and a video-based skeleton baseline

Load-bearing premise

The claim that wearable sensing can replace cameras in real hazardous conditions rests on the assumption that six indoor participants adequately represent the range of operators and environments; all data were collected indoors, so the robustness advantage in smoke, low light, and clutter is asserted rather than demonstrated.

What would settle it

Record the same 20 gestures with the same sensors and a camera in smoke-filled and low-light settings with the full eleven participants; if the camera baseline's F1 does not fall below the LLR sensor model's, or if the sensor model degrades just as much, the paper's central robustness-and-parity claim is falsified.

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

If this is right

  • A sensor-only setup—no camera, no line of sight—matches a video-based skeleton model on F1 for the 20-gesture vocabulary, with differences of roughly one percentage point or less.
  • The wearable model needs far fewer GFLOPs per inference, a smaller model file, and shorter training time, so real-time gesture control could run on edge devices rather than in the cloud.
  • Because LLR fusion assigns an explicit signed contribution to each sensor stream, a designer can trace a mistaken command to the sensors that drove it, aiding debugging and operator trust.
  • The approach is intended to keep working where camera-based gesture recognition fails—smoke, low light, occlusion—since inertial and capacitive signals do not depend on visible light; this is the stated motivation, not a tested condition in the paper.
  • The new synchronized RGB-IMU-capacitive dataset gives the community a shared 20-gesture teleoperation benchmark grounded in the aircraft marshalling protocol.

Where Pith is reading between the lines

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

  • The natural next experiment is to record the same gestures under fog, darkness, and outdoor clutter with the full participant set; if camera accuracy drops while the sensor model stays high, the paper's central robustness claim is directly confirmed.
  • The LLR fusion rule is a generic late-fusion layer: it can be dropped into any multi-stream classifier, and its log-odds outputs could be thresholded for uncertain-command rejection or converted to 'soft' control signals, which the paper does not explore.
  • The ablation's finding that capacitive sensing adds little for large arm motions suggests the sensor set is underutilized by the current vocabulary; adding finger-level gestures like pinches or taps would likely rebalance modality use and strengthen the multimodal argument.
  • Since the reported F1 differences between the sensor model and the vision baseline are within one standard deviation, the more durable claim is the efficiency-plus-interpretability package; an online closed-loop teleoperation demo with latency measurements would test the real-time suitability claim directly.

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

4 major / 6 minor

Summary. The paper proposes a wearable-sensor gesture recognition framework for drone and mobile robot teleoperation, combining inertial measurements from two Apple Watches with capacitive sensing from custom gloves, and fusing per-modality classifiers with a log-likelihood ratio (LLR) late-fusion rule. A new dataset of 20 aircraft-marshalling-inspired gestures with synchronized RGB, IMU, and capacitive data is introduced. The authors evaluate LLR fusion and self-attention fusion under LOSO and LOPO splits, compare against PoseConv3D as a vision baseline, and report that the sensor-based approach achieves comparable or better F1 scores while requiring substantially fewer computational resources. Ablation studies and interpretability analyses aim to characterize the contribution of each modality and the interpretability of the fusion weights.

Significance. If the central comparison is substantiated, the paper would make a useful practical contribution: a low-cost, wearable-only gesture recognition system matching a strong vision-based method for teleoperation would be valuable for real-time robot control, especially in smoke, low light, or cluttered environments where vision degrades. The introduction of a synchronized multimodal dataset is a tangible community resource, and the LLR fusion rule has the merit of being a simple, parameter-free, late-fusion operation with post-hoc interpretability that the authors explicitly hedge rather than overclaim. The code and dataset release are also a strength. However, the manuscript's central quantitative claim—that the sensor model is comparable to or better than PoseConv3D—is not yet supported by appropriate statistical evidence, and the practical robustness claims go beyond the indoor, six-participant evaluation actually conducted.

major comments (4)
  1. [§V-A, Table IV] The claim that the sensor-based LLR model is 'comparable' to or 'outperforms' PoseConv3D is not statistically supported. The F1 differences are 0.70 (LOSO) and 0.20 (LOPO), and the standard deviations overlap substantially (e.g., 93.59±5.20 vs. 93.39±5.44 for LOPO). With LOPO evaluated on only six participants, the data are consistent with the sensor model being materially worse. The paper reports no paired significance test, no confidence intervals for the difference, and no pre-specified equivalence margin. 'Comparable' is an equivalence claim and needs an equivalence test or at least a confidence-interval argument. This is load-bearing because the entire resource-efficiency argument depends on the sensor model being no worse than the vision baseline. Additionally, the Introduction states the approach 'outperforms' vision, which is stronger than the abstract's 'comparable' and should b
  2. [§IV-C and §V] The LOPO evaluation uses only six participants, after excluding corrupted sessions, and all data were collected indoors. The paper nevertheless makes robustness claims in the Discussion about smoke, low light, and occlusion. Those claims are not tested by the present experiments. At minimum, the authors should state explicitly that these are motivating scenarios, not demonstrated results, and should temper the generality of the conclusions. More importantly, with n=6, the LOPO standard deviations of about five F1 points mean that per-participant variability dominates the comparison; reporting participant-level results or confidence intervals per split would help the reader assess how stable the observed F1 values are.
  3. [§V-C, Table V] The ablation results undermine the 'multimodal' claim as presented. The capacitive-only model collapses to a single class in the LOPO split (F1=7.98±4.27), and the full four-sensor model (93.59 LOPO) does not outperform the three-sensor ACC+GYRO+QUAT configuration (93.75 LOPO). The Discussion acknowledges that capacitive sensing contributed minimally, but this is a major point: the added hardware complexity of the custom gloves is not justified by the reported data. The claim that capacitive signals provide 'complementary information' is supported only qualitatively, not by the quantitative results. The authors should either provide a more controlled analysis showing when capacitive sensing adds value or substantially downgrade this contribution in the framing.
  4. [§IV-C and §V-A] The PoseConv3D baseline is under-specified. PoseConv3D is a skeleton-based action recognition method, but the paper does not state which pose estimator was used, how skeletons were extracted from the RGB video, what preprocessing was applied, or which hyperparameters were selected. Without these details, the comparison is not auditable, and the computational resource comparison in Fig. 5 may reflect implementation choices rather than inherent properties of the two approaches. Please provide a complete baseline description, including the skeleton extraction pipeline, the training protocol, and any hyperparameter tuning, or otherwise clarify the level of fair comparison.
minor comments (6)
  1. [Eq. (1)] The notation is slightly inconsistent: 'N classes {C_i, ..., C_n}' mixes N and n. Please use one index convention consistently.
  2. [Fig. 2 caption] The caption is confusing: 'Purple outlines mark capacitive sensors (bottom) and IMU sensors (top)'—it is unclear whether the purple outlines mark both the bottom capacitive sensors and top IMU sensors, or whether these are separate elements. Please clarify.
  3. [§IV-B] The temporal scaling factor is introduced as the ratio of video to sensor sequence length, but it is not described how it is computed in practice or how robust it is to synchronization errors. Please give one concrete example.
  4. [§IV-B] The label threshold of 0.75 is reported but not justified. Since windows below the threshold are discarded, this choice affects the effective training set size and could influence the reported performance. A brief sensitivity check would improve confidence.
  5. [§V-B, Figs. 6 and 7] The interpretability analysis is based on 'randomly sampled' examples for a single gesture. This is anecdotal. If the goal is to quantify modality contributions, an aggregate analysis over all correctly classified test windows (e.g., mean absolute LLR per class) would be more informative and less prone to cherry-picking impressions.
  6. [§V-D] The Discussion says 'the glove sensors provided complementary information in cases where wrist-based inertial signals were ambiguous,' but no evidence or quantified example is given. Either add such evidence or soften this statement.

Circularity Check

0 steps flagged

No significant circularity: LLR fusion is a parameter-free late-fusion rule and all performance claims are empirical, not derived from fitted inputs.

full rationale

The paper's central derivation chain is the LLR fusion rule: Eq. (1) defines the per-modality log-likelihood ratio from the modality's own softmax output, and Eq. (2) sums these LLRs across the eight modality pairs. This fusion rule contains no parameters fitted to the target accuracy figure; the only learned quantities are the per-modality feature extractors and the final linear projection, trained with cross-entropy on held-out LOSO/LOPO splits. The 'predictions' that are compared to PoseConv3D in Table IV are empirical evaluation results, not quantities derived from a fitted input by construction. The interpretability analysis (Figs. 6-7) is post-hoc: LLR values and attention weights are computed from an already-trained model and are explicitly hedged ('we treat them as diagnostic, as high attention does not guarantee a causal role in the final prediction'). Self-citations ([12], [13], [28]) supply the CAPTAINGlove hardware platform and prior HAR context, but do not themselves produce the reported F1 scores; the gloves are re-evaluated on a new dataset in this paper. No uniqueness theorem, ansatz, or fitted parameter is imported via self-citation. The 'comparable to vision' claim is statistically fragile (n=6 LOPO, overlapping standard deviations, no equivalence test), but that is an evidence-strength issue, not a circular derivation.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The paper introduces no new entities. It relies on learned network weights and a set of hand-chosen hyperparameters, plus domain assumptions about sensor synchronization and generalization from indoor data to hazardous environments.

free parameters (6)
  • window_size = 3.0 s
    Chosen by hand to capture temporal context; directly affects all results.
  • step_size = 1.0 s
    Sliding window step; hand-chosen, influences number of training windows.
  • label_threshold = 0.75
    Threshold for window label voting; determines which windows are kept or discarded.
  • temporal_scaling_factor = per-session ratio
    Computed from clap synchronization to stretch/compress sensor data to match video; affects cross-modal alignment.
  • learning_rate = 1e-4
    Adam hyperparameter; hand-chosen.
  • batch_size = 32
    Training hyperparameter; hand-chosen.
axioms (4)
  • domain assumption Clap-based synchronization and temporal scaling produce accurate cross-modal alignment for all gesture windows.
    Section IV-B: five claps at start/end of each session and a scaling factor are used to align video and sensor streams; if misaligned, window labels are wrong.
  • domain assumption Per-modality softmax outputs can be treated as independent class likelihoods and fused by summing LLRs (naive independence).
    Eq. (1)-(2): the fused representation is a sum of per-modality LLRs, which assumes conditional independence across the eight modality inputs.
  • domain assumption The 20 aircraft-marshalling gestures are distinct, performable with natural variation, and captured by the sensor set.
    Section IV-A defines the gesture set; the ablation shows capacitive sensors fail on most gestures, challenging this assumption for that modality.
  • domain assumption Indoor-collected data generalizes to hazardous environments such as smoke-filled rooms and low-light industrial sites.
    Discussion claims robustness in smoke/low light while acknowledging all data was indoors (Section V-D).

pith-pipeline@v1.3.0-alltime-deepseek · 11552 in / 8093 out tokens · 79033 ms · 2026-08-02T20:10:43.649713+00:00 · methodology

0 comments
read the original abstract

Human operators are still frequently exposed to hazardous environments such as disaster zones and industrial facilities, where intuitive and reliable teleoperation of mobile robots and Unmanned Aerial Vehicles (UAVs) is essential. In this context, hands-free teleoperation enhances operator mobility and situational awareness, thereby improving safety in hazardous environments. While vision-based gesture recognition has been explored as one method for hands-free teleoperation, its performance often deteriorates under occlusions, lighting variations, and cluttered backgrounds, limiting its applicability in real-world operations. To overcome these limitations, we propose a multimodal gesture recognition framework that integrates inertial data (accelerometer, gyroscope, and orientation) from Apple Watches on both wrists with capacitive sensing signals from custom gloves. We design a late fusion strategy based on the log-likelihood ratio (LLR), which not only enhances recognition performance but also provides interpretability by quantifying modality-specific contributions. To support this research, we introduce a new dataset of 20 distinct gestures inspired by aircraft marshalling signals, comprising synchronized RGB video, IMU, and capacitive sensor data. Experimental results demonstrate that our framework achieves performance comparable to a state-of-the-art vision-based baseline while significantly reducing computational cost, model size, and training time, making it well suited for real-time robot control. We therefore underscore the potential of sensor-based multimodal fusion as a robust and interpretable solution for gesture-driven mobile robot and drone teleoperation.

Figures

Figures reproduced from arXiv: 2602.23694 by Hymalai Bello, Jaspreet Singh, Lala Shakti Swarup Ray, Paul Lukowicz, Seungyeol Baek, Sungho Suh.

Figure 1
Figure 1. Figure 1: An example of drone control using the proposed [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Hardware setup with textile sensing gloves and Apple [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Framework architecture for sensor-based gesture [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Example gesture classes are illustrated, including [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Computational resource comparison between sensor [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Randomly sampled LLR contribution per modality, [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Randomly sampled attention weights for a case when [PITH_FULL_IMAGE:figures/full_fig_p006_7.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

39 extracted references · 5 linked inside Pith

  1. [1]

    An android based mobile robot for monitoring and surveillance,

    J. Azeta, C. Bolu, D. Hinvi, A. Abioye, H. Boyo, P. Anakhu, and P. Onwordi, “An android based mobile robot for monitoring and surveillance,”Procedia Manufacturing, vol. 35, pp. 1129–1134, 2019

    2019

  2. [2]

    Real-time monitoring of work-at-height safety hazards in construction sites using drones and deep learning,

    M. Z. Shanti, C.-S. Cho, B. G. de Soto, Y .-J. Byon, C. Y . Yeun, and T. Y . Kim, “Real-time monitoring of work-at-height safety hazards in construction sites using drones and deep learning,”Journal of safety research, vol. 83, pp. 364–370, 2022

    2022

  3. [3]

    Training for smart manufacturing using a mobile robot-based production line,

    S. Wang, L. Jiang, J. Meng, Y . Xie, and H. Ding, “Training for smart manufacturing using a mobile robot-based production line,”Frontiers of Mechanical Engineering, vol. 16, no. 2, pp. 249–270, 2021

    2021

  4. [4]

    Do drones have a realistic place in a pandemic fight for delivering medical supplies in healthcare systems problems?

    J. Euchi, “Do drones have a realistic place in a pandemic fight for delivering medical supplies in healthcare systems problems?” pp. 182– 190, 2021

    2021

  5. [5]

    A review of mobile robots: Concepts, methods, theoretical framework, and applications,

    F. Rubio, F. Valero, and C. Llopis-Albert, “A review of mobile robots: Concepts, methods, theoretical framework, and applications,” International Journal of Advanced Robotic Systems, vol. 16, no. 2, p. 1729881419839596, 2019

    2019

  6. [6]

    Emerging uav technology for disaster detection, mitigation, response, and preparedness,

    A. Khan, S. Gupta, and S. K. Gupta, “Emerging uav technology for disaster detection, mitigation, response, and preparedness,”Journal of Field Robotics, vol. 39, no. 6, pp. 905–955, 2022

    2022

  7. [7]

    Vehicle teleoperation interfaces,

    T. Fong and C. Thorpe, “Vehicle teleoperation interfaces,”Autonomous robots, vol. 11, no. 1, pp. 9–18, 2001

    2001

  8. [8]

    A mixed-reality tele-operation method for high-level control of a legged- manipulator robot,

    C. Cruz Ulloa, D. Dom ´ınguez, J. Del Cerro, and A. Barrientos, “A mixed-reality tele-operation method for high-level control of a legged- manipulator robot,”Sensors, vol. 22, no. 21, p. 8146, 2022

    2022

  9. [9]

    Gestures-teleoperation of a heterogeneous multi-robot system,

    K. B. de Carvalho, D. K. D. Villa, M. Sarcinelli-Filho, and A. S. Bran- dao, “Gestures-teleoperation of a heterogeneous multi-robot system,” The International Journal of Advanced Manufacturing Technology, vol. 118, no. 5, pp. 1999–2015, 2022

    1999

  10. [10]

    Hand gesture recognition with convolutional neural networks for the multimodal uav control,

    Y . Ma, Y . Liu, R. Jin, X. Yuan, R. Sekha, S. Wilson, and R. Vaidyanathan, “Hand gesture recognition with convolutional neural networks for the multimodal uav control,” in2017 Workshop on Research, Education and Development of Unmanned Aerial Systems (RED-UAS). IEEE, 2017, pp. 198–203

    2017

  11. [11]

    Vision based hand gesture recognition,

    Y . Zhu, Z. Yang, and B. Yuan, “Vision based hand gesture recognition,” in2013 international conference on service sciences (ICSS). IEEE, 2013, pp. 260–265

    2013

  12. [12]

    Worker activity recognition in manufacturing line using near-body electric field,

    S. Suh, V . F. Rey, S. Bian, Y .-C. Huang, J. M. Ro ˇzanec, H. T. Ghinani, B. Zhou, and P. Lukowicz, “Worker activity recognition in manufacturing line using near-body electric field,”IEEE Internet of Things Journal, vol. 11, no. 7, pp. 11 554–11 565, 2023

    2023

  13. [13]

    Captainglove: Capacitive and inertial fusion-based glove for real- time on edge hand gesture recognition for drone control,

    H. Bello, S. Suh, D. Geißler, L. S. S. Ray, B. Zhou, and P. Lukowicz, “Captainglove: Capacitive and inertial fusion-based glove for real- time on edge hand gesture recognition for drone control,” inAdjunct Proceedings of the 2023 ACM International Joint Conference on Pervasive and Ubiquitous Computing & the 2023 ACM International Symposium on Wearable Com...

    2023

  14. [14]

    Visual recognition of aircraft marshalling signals using gesture phase analysis,

    C. Choi, J.-H. Ahn, and H. Byun, “Visual recognition of aircraft marshalling signals using gesture phase analysis,” in2008 IEEE Intelligent Vehicles Symposium. IEEE, 2008, pp. 853–858

    2008

  15. [15]

    Revisiting skeleton- based action recognition,

    H. Duan, Y . Zhao, K. Chen, D. Lin, and B. Dai, “Revisiting skeleton- based action recognition,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 2969–2978

    2022

  16. [16]

    Tele- operation methods and enhancement techniques for mobile robots: A comprehensive survey,

    M. Moniruzzaman, A. Rassau, D. Chai, and S. M. S. Islam, “Tele- operation methods and enhancement techniques for mobile robots: A comprehensive survey,”Robotics and Autonomous Systems, vol. 150, p. 103973, 2022

    2022

  17. [17]

    A review on teleoperation of mobile ground robots: Architecture and situation awareness,

    S. Opiyo, J. Zhou, E. Mwangi, W. Kai, and I. Sunusi, “A review on teleoperation of mobile ground robots: Architecture and situation awareness,”International Journal of Control, Automation and Systems, vol. 19, no. 3, pp. 1384–1407, 2021

    2021

  18. [18]

    Mixed-initiative variable autonomy for remotely operated mobile robots,

    M. Chiou, N. Hawes, and R. Stolkin, “Mixed-initiative variable autonomy for remotely operated mobile robots,”ACM Transactions on Human-Robot Interaction (THRI), vol. 10, no. 4, pp. 1–34, 2021

    2021

  19. [19]

    Teleman: Teleoperation for legged robot loco-manipulation using wearable imu- based motion capture,

    C. Zhou, C. Peers, Y . Wan, R. Richardson, and D. Kanoulas, “Teleman: Teleoperation for legged robot loco-manipulation using wearable imu- based motion capture,”arXiv preprint arXiv:2209.10314, 2022

    Pith/arXiv arXiv 2022

  20. [20]

    Immersive robot teleoperation based on user gestures in mixed reality space,

    H. Esaki and K. Sekiyama, “Immersive robot teleoperation based on user gestures in mixed reality space,”Sensors, vol. 24, no. 15, p. 5073, 2024

    2024

  21. [21]

    Investigating the usability of collabo- rative robot control through hands-free operation using eye gaze and augmented reality,

    J. Lee, T. Lim, and W. Kim, “Investigating the usability of collabo- rative robot control through hands-free operation using eye gaze and augmented reality,” in2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023, pp. 4101–4106

    2023

  22. [22]

    Analysis of mr–vr tele-operation methods for legged-manipulator robots,

    C. Cruz Ulloa, D. Dom ´ınguez, J. del Cerro, and A. Barrientos, “Analysis of mr–vr tele-operation methods for legged-manipulator robots,”Virtual Reality, vol. 28, no. 3, p. 131, 2024

    2024

  23. [23]

    Horus: A mixed reality interface for managing teams of mobile robots,

    O. S. Adekoya, A. Sgorbissa, and C. T. Recchiuto, “Horus: A mixed reality interface for managing teams of mobile robots,”arXiv preprint arXiv:2506.02622, 2025

    Pith/arXiv arXiv 2025

  24. [24]

    Toward intuitive 3d interactions in virtual reality: A deep learning-based dual-hand gesture recognition approach,

    T. Di Qi, F. L. Cibrian, M. Raswan, T. Kay, H. M. Camarillo-Abad, and Y . Wen, “Toward intuitive 3d interactions in virtual reality: A deep learning-based dual-hand gesture recognition approach,”IEEE Access, vol. 12, pp. 67 438–67 452, 2024

    2024

  25. [25]

    Hand gesture recognition gaming control system: Harnessing hand gestures and voice commands for immersive gameplay,

    A. Sharma, L. Verma, H. Kaur, A. Modgil, A. Soniet al., “Hand gesture recognition gaming control system: Harnessing hand gestures and voice commands for immersive gameplay,” in2024 International Conference on Emerging Innovations and Advanced Computing (IN- NOCOMP). IEEE, 2024, pp. 101–107

    2024

  26. [26]

    Multidisciplinary ml techniques on gesture recognition for people with disabilities in a smart home environment,

    C. Panagiotou, E. Faliagka, C. P. Antonopoulos, and N. V oros, “Multidisciplinary ml techniques on gesture recognition for people with disabilities in a smart home environment,”AI, vol. 6, no. 1, p. 17, 2025

    2025

  27. [27]

    Computer vision-based hand gesture recognition for human-robot interaction: a review,

    J. Qi, L. Ma, Z. Cui, and Y . Yu, “Computer vision-based hand gesture recognition for human-robot interaction: a review,”Complex & Intelligent Systems, vol. 10, no. 1, pp. 1581–1606, 2024

    2024

  28. [28]

    Als-har: Harnessing wearable ambient light sensors to enhance imu- based human activity recognition,

    L. S. S. Ray, D. Geißler, M. Liu, B. Zhou, S. Suh, and P. Lukowicz, “Als-har: Harnessing wearable ambient light sensors to enhance imu- based human activity recognition,” inInternational Conference on Pattern Recognition. Springer, 2024, pp. 133–147

    2024

  29. [29]

    Empirical evaluation of gated recurrent neural networks on sequence modeling,

    J. Chung, C. Gulcehre, K. Cho, and Y . Bengio, “Empirical evaluation of gated recurrent neural networks on sequence modeling,”arXiv preprint arXiv:1412.3555, 2014

    Pith/arXiv arXiv 2014

  30. [30]

    Deep convolutional and lstm recur- rent neural networks for multimodal wearable activity recognition,

    F. J. Ord ´o˜nez and D. Roggen, “Deep convolutional and lstm recur- rent neural networks for multimodal wearable activity recognition,” Sensors, vol. 16, no. 1, p. 115, 2016

    2016

  31. [31]

    Tinyhar: A lightweight deep learning model designed for human activity recognition,

    Y . Zhou, H. Zhao, Y . Huang, T. Riedel, M. Hefenbrock, and M. Beigl, “Tinyhar: A lightweight deep learning model designed for human activity recognition,” inProceedings of the 2022 ACM International Symposium on Wearable Computers, 2022, pp. 89–93

    2022

  32. [32]

    Attnsense: Multi-level attention mechanism for multimodal human activity recognition

    H. Ma, W. Li, X. Zhang, S. Gao, and S. Lu, “Attnsense: Multi-level attention mechanism for multimodal human activity recognition.” in IJCAI, 2019, pp. 3109–3115

    2019

  33. [33]

    Attend and discriminate: Beyond the state-of-the-art for human activity recognition using wearable sensors,

    A. Abedin, M. Ehsanpour, Q. Shi, H. Rezatofighi, and D. C. Ranas- inghe, “Attend and discriminate: Beyond the state-of-the-art for human activity recognition using wearable sensors,”Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 5, no. 1, pp. 1–22, 2021

    2021

  34. [34]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems, vol. 30, 2017

    2017

  35. [35]

    ” why should i trust you?

    M. T. Ribeiro, S. Singh, and C. Guestrin, “” why should i trust you?” explaining the predictions of any classifier,” inProceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, 2016, pp. 1135–1144

    2016

  36. [36]

    Is attention interpretable?

    S. Serrano and N. A. Smith, “Is attention interpretable?”arXiv preprint arXiv:1906.03731, 2019

    Pith/arXiv arXiv 1906

  37. [37]

    Label Studio: Data labeling software,

    M. Tkachenko, M. Malyuk, A. Holmanyuk, and N. Liubimov, “Label Studio: Data labeling software,” 2020-2025, open source software available from https://github.com/HumanSignal/label-studio. [Online]. Available: https://github.com/HumanSignal/label-studio

    2020

  38. [38]

    Adam: A method for stochastic optimiza- tion,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimiza- tion,”arXiv preprint arXiv:1412.6980, 2014

    Pith/arXiv arXiv 2014

  39. [39]

    Cellular-enabled collaborative robots planning and operations for search-and-rescue scenarios,

    A. Romero, C. Delgado, L. Zanzi, R. Su ´arez, and X. Costa-P ´erez, “Cellular-enabled collaborative robots planning and operations for search-and-rescue scenarios,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 5942–5948

    2024