Pith. sign in

REVIEW 3 major objections 6 minor 39 references

ARGO shows a smart-eyewear camera can run YOLOv11 on-device at 10 FPS with a 2.483 MB model and about 113 minutes of battery life.

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 10:52 UTC pith:NXYPAZEJ

load-bearing objection Real platform, credible measurements; the headline mAP is optimistic because the same test set was used for model selection and the evaluator allows 300 detections while the shipped pipeline caps at 20. the 3 major comments →

arxiv 2607.16222 v1 pith:NXYPAZEJ submitted 2026-06-17 cs.LG cs.AIeess.SP

Fully-sensorized smart-eyewear platform for on-device Machine Learning

Andrea Giudici , Christian Veronesi , Pietro Bartoli , Mario Cali\`o , Aurelio Teliti , Giacomo Gervasoni , Diana Trojaniello , Franco Zappa This is my paper
classification cs.LG cs.AIeess.SP
keywords smart eyewearon-device machine learningTinyMLYOLOv11STM32N6 NPUHead-wise Parallel Attentionegocentric visionurban obstacle detection
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 tries to establish that a comfortable smart-eyewear device can run a modern object detector entirely on its own embedded neural accelerator, without cloud offloading. The authors build ARGO around the STM32N6 microcontroller's NPU, co-designing the hardware, firmware, and the YOLOv11 model so that detection runs at 10 FPS. They introduce Head-wise Parallel Attention (HPA), a rewrite of the attention block that uses only 3D matrix multiplications the NPU supports, with no change to the attention logic. The deployed model uses only 2.483 MB and achieves 0.24 mAP50-95 on the WOTR pedestrian-obstacle dataset after INT8 quantization, while the whole platform draws about 113 minutes of continuous operation from a 200 mAh battery. If true, this shows that privacy-preserving, real-time egocentric vision is feasible within eyewear ergonomics.

Core claim

On its own terms, the paper's central claim is that the gap between smart-eyewear ergonomics and real-time vision AI can be closed by simultaneous co-design of the electronics, the firmware pipeline, and the neural network. The platform's STM32N6 NPU executes YOLOv11n at 256x256 input with 143 of 170 inference epochs on the NPU and all activations in internal SRAM, consuming a 2.483 MB weight footprint. After symmetric per-channel INT8 quantization of weights and percentile-calibrated asymmetric activation quantization, the detector reaches 0.24 mAP50-95 on the WOTR test set (0.28 in FP32), with host-device output agreement at cosine similarity 0.9996. The full capture-process-infer-postproc

What carries the argument

HPA (Head-wise Parallel Attention) is the central mechanism: it restructures multi-head attention so that the usual batched-head 4D matrix multiplications become H independent 3D matmuls, one per head, concatenated afterward. This lets YOLOv11's attention block compile onto the NPU, which only supports 3D tensor kernels, instead of failing or falling back to CPU execution; the attention function itself is preserved. The other load-bearing mechanism is the deployment constraint that keeps all activation memory internal to the MCU and the weight footprint within non-volatile storage, which selects YOLOv11n at 256x256 as the operating point.

Load-bearing premise

The detection accuracy claim depends on WOTR, a dataset captured from a non-head-mounted perspective, being representative of what ARGO's camera actually sees when worn by a pedestrian; no egocentric test set is used to verify this.

What would settle it

Collect a few hundred egocentric frames from the actual glasses worn by pedestrians, annotate them with the same 20 classes, run the deployed YOLOv11n-256 model, and compare mAP50-95 to the 0.24 reported on WOTR; a large drop or a systematic shift in object scale or viewpoint would falsify the representativeness assumption.

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

If this is right

  • On-device detection runs at low latency with no raw image leaving the device, since only detection metadata is transmitted over BLE.
  • The same co-design template can be applied to other NPU-equipped MCUs and other attention-based vision models once their attention blocks are rewritten in HPA style.
  • INT8 post-training quantization costs are measurable and scale-sensitive, with relative mAP drops up to about 25% for larger variants, so accuracy must be reported together with memory and energy constraints.
  • The priority-based truncation to at most 20 detections per frame keeps the over-the-air data rate within BLE bandwidth even though the network emits up to 2100 raw predictions per frame.

Where Pith is reading between the lines

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

  • If the WOTR-to-egocentric transfer holds in field use, similar single-chip NPU glasses could run continuous obstacle detection for a full commute; the paper leaves that confirmation to future egocentric benchmarks.
  • The priority-based truncation is a safety-motivated lossy step that the paper tunes but does not evaluate against missed-detection rates; a natural follow-up is to measure how many true high-risk objects are dropped when the 20-detection budget is enforced.
  • Because the HPA change is claimed to be evaluation-order-only, it suggests a general recipe: any attention block can be made NPU-friendly by expressing heads as separate 3D gemms, a testable claim for other detectors on the same class of accelerators.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. ARGO is a smart-eyewear platform built around the STM32N6 MCU and its integrated NPU, with a multimodal sensor suite. The paper's central deployment claims are that a YOLOv11n detector at 256×256 input runs on-device at 10 FPS with a 2.483 MB weight footprint and fully internal activations, reaches mAP50–95 = 0.24 after INT8 PTQ on the WOTR test set, and provides roughly 113 minutes of autonomy from a 200 mAh battery. To make the model NPU-compatible, the authors introduce Head-wise Parallel Attention (HPA), a rearrangement of multi-head attention into 3D matrix multiplications that is algebraically equivalent to the original formulation. The paper reports a systematic model-selection grid over three YOLO variants and four input resolutions, detailed power/energy measurements per subsystem, and a full battery discharge characterization.

Significance. If the accuracy and autonomy claims hold, this is a useful demonstration of end-to-end co-design for on-device machine learning in smart eyewear: the HPA reformulation is mathematically transparent and potentially reusable for other NPUs; the power-measurement methodology is careful; and the on-device versus host-side numerical agreement over the full test set is a genuine strength. The systems contribution—10 FPS on a 200 mAh form factor—is significant for the wearable edge-AI community. However, the headline accuracy figure is not yet a valid estimate of deployed performance, for the reasons detailed below, so the feasibility claim currently rests on weaker evidence than the system measurements.

major comments (3)
  1. [§5.1, Table 2] The reported mAP50–95 = 0.24 is computed on the WOTR test split, but this same split was used to select the deployed configuration: Section 5.1 states that the full n/s/m × {128,192,256,320} grid was benchmarked on the WOTR test set, and the best eligible operating point was then reported on that same set. Selecting the best of 12 configurations on the test set and then reporting that configuration's test-set accuracy introduces optimistic selection bias. The headline accuracy should be reported on a separate held-out set, or model selection should be performed on a validation split and the test set used only once. This issue is load-bearing because mAP = 0.24 is the central accuracy result.
  2. [§4.3.1] The only bridge between the reported mAP and ARGO's deployed egocentric camera is the assertion that WOTR, although 'not captured directly from a head-mounted device,' has visual content that 'closely reflects what a pedestrian typically observes.' No egocentric validation set is used anywhere in the paper, so the effects of head-mounted camera height, pitch, ego-motion, and occlusion patterns on detection accuracy are unmeasured. Since the stated application is assistive navigation, this domain-transfer assumption is correctness-critical. A concrete fix is to collect and annotate a small egocentric test set with ARGO itself and report mAP on it, or at minimum provide quantitative evidence of viewpoint similarity (e.g., object-scale and truncation statistics). Without this, the deployed model's real-world performance is unknown.
  3. [§4.4] The evaluation protocol is inconsistent with the deployed post-processing budget. The paper states that evaluation uses 'a maximum of 300 detections per image,' while the deployed on-device pipeline truncates detections to Nmax = 20 after priority scoring and before NMS. The reported mAP therefore corresponds to a substantially looser detection budget than the one actually used by ARGO when worn. The authors should recompute all reported metrics using the deployed priority-scoring and Nmax = 20 truncation, or explicitly report both numbers. The current wording, 'optional truncation to a fixed detection budget,' does not remedy the mismatch, because the stated evaluation setting explicitly avoids the stricter budget.
minor comments (6)
  1. [Table 3] Several entries in Table 3 appear corrupted in the submitted text, e.g., '40.79±0.010.76 0.88 0.168.05±0.1067.69 68.35 0.645.21±0.0744.92 5.29 0.36.' Please reformat the table so that mean, standard deviation, min, max, and peak-to-peak values are legible.
  2. [Figure 10] The axis labels and panel annotations in Figure 10 are rendered as garbled glyph sequences (e.g., '/uni00000015 /uni00000017'). The figure should be regenerated with a standard font encoding so the power traces can be interpreted.
  3. [Abstract] The abstract reports 'mAP50–95 of 24' while Table 2 uses decimal notation (0.24). Use consistent units (e.g., '0.24' or '24%') throughout to avoid ambiguity.
  4. [§4.4] The confidence threshold τ_conf used in evaluation is described only as 'low.' Please give the exact value so the metric is reproducible.
  5. [§5.1, Fig. 8] The qualitative detection examples are said to show the 'egocentric viewpoint of the eyewear,' but it is not clear whether these frames come from the WOTR test set or from ARGO's own camera. Please state the source of the images in Figure 8.
  6. [§4.3.2] The HPA derivation (Eqs. 3–5) is mathematically a rearrangement of standard multi-head attention, and the claim that it introduces no modeling error beyond floating-point evaluation order is plausible. However, the paper should state clearly whether the models were trained with the original MHA and converted to HPA at deployment, or trained directly in HPA form; this affects reproducibility.

Circularity Check

0 steps flagged

No significant circularity: central claims rest on direct measurements and a self-contained algebraic reformulation; only a minor non-load-bearing self-citation is present.

full rationale

The paper's principal claims are empirical or algebraic rather than derivations from fitted constants. The HPA reformulation in Section 4.3.2 (Eqs. 3-5) is a tensor-rank rearrangement of standard multi-head attention: each head computes S_i = Q_i^T K_i and O_i = V_i A_i^T, which is exactly the per-head matrix multiplication of the original 4D MHA. The statement that the reformulation 'introduces no additional modeling error beyond floating-point evaluation order' is a mathematical identity, not a circular definition. The accuracy, memory, latency, and power figures (Tables 2-3, Figs. 7, 10-11) are measured on the WOTR test set and on the ARGO hardware, not predicted from the paper's own assumptions. The self-citation to [38] for the power-benchmarking methodology is minor and non-load-bearing: the measurement procedure (shunt resistors, differential probes, oscilloscope triggering) is described in the paper itself and is standard instrumentation practice. Several other same-group references ([7], [8], [30], [32]) are contextual and do not support the central derivation. The main validity caveats—WOTR is not egocentric (Section 4.3.1), the WOTR test set was used for model selection (Section 5.1), and evaluation uses up to 300 detections while deployed post-processing truncates to 20 (Section 4.4)—are statistical/domain-transfer limitations, not circularity: the paper does not define its reported mAP in terms of these assumptions and explicitly discloses the detection-budget discrepancy. Overall, the derivation chain is self-contained enough that no prediction reduces to its inputs by construction.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entities. Its free parameters are mostly hand-chosen engineering settings; the most significant is the test-set-selected model configuration, which affects the reported accuracy. The key domain assumptions concern the representativeness of WOTR and the NPU/hardware behavior, neither of which is independently validated here.

free parameters (6)
  • Evaluation confidence threshold (τ_conf) = not reported
    mAP is computed after filtering with an unspecified 'low' confidence threshold; this threshold affects the reported mAP and is not quantified.
  • NMS IoU threshold (τ_IoU) = 0.6
    Chosen by hand in §4.4; affects the final detections and therefore the measured mAP.
  • Activation quantization percentile = 99.999%
    Clipping tails at 99.999% during PTQ calibration (§4.3.3) affects INT8 accuracy; chosen by hand.
  • Deployed model configuration (YOLOv11n, input 256) = weights 2.483 MB, act_int 632 kB
    Selected after evaluating the full grid on the WOTR test set (Table 2); the reported mAP is for this selected configuration and may be optimistic due to test-set selection.
  • CPU/NPU/RAM clock frequency = 600 MHz
    Chosen as a trade-off between FPS and power (§5.2); the 10 FPS and 113 min autonomy depend on this value.
  • Priority scoring parameters (α, P[cls]) = α=0.1, P=1–5
    Hand-set for deployed post-processing (§4.4); not used in the reported evaluation, so they do not affect the mAP claim but do affect the deployed system's behavior.
axioms (5)
  • standard math HPA's sequential 3D matmuls are mathematically equivalent to the original 4D multi-head attention (up to transpose and concatenation).
    Eqs. (3)–(5) and surrounding text in §4.3.2; this is a reshape/view of the same computations.
  • domain assumption The WOTR dataset is representative of egocentric views from a head-mounted device.
    §4.3.1 states 'WOTR is not captured directly from a head-mounted device, its visual content closely reflects...'; no egocentric validation set is provided.
  • domain assumption The STM32N6 NPU executes the compiled graph with operator support and epoch scheduling as described in ST documentation [37].
    §4.3.2 relies on the NPU's 3D-matmul support and hybrid execution behavior; not independently verified in this paper.
  • domain assumption The power measurement methodology of [38] (shunt resistors, synchronized oscilloscopes) yields accurate rail current/energy values.
    §5.2 adopts the benchmarking method from the authors' previous work [38] without independent validation.
  • domain assumption Post-training quantization calibration statistics from the WOTR train set generalize to deployment scenes.
    §4.3.3 uses a percentile calibration on the training set; assumes the distribution matches real-world ARGO scenes.

pith-pipeline@v1.3.0-alltime-deepseek · 21195 in / 18599 out tokens · 164943 ms · 2026-08-02T10:52:50.854190+00:00 · methodology

0 comments
read the original abstract

This paper presents ARGO, a smart eyewear platform designed to bridge ergonomic comfort, high computational throughput, and energy efficiency. Unlike cloud-dependent solutions, ARGO leverages the STM32N6 microcontroller and its integrated Neural Processing Unit (NPU) to enable on-device machine learning, minimizing latency and preserving user privacy through local data processing. The primary contribution lies in the holistic co-design of hardware, firmware, and artificial intelligence, centered on the deployment of an optimized YOLOv11 model for real-time urban obstacle recognition. To ensure compatibility with the target NPU, we introduce Head-wise Parallel Attention (HPA), an architectural refinement that enables efficient accelerator execution while preserving the original computational logic. The model is trained on the Walking On The Road (WOTR) dataset, and the final deployed configuration achieves an mAP50-95 of 24 under strict memory constraints, with a memory footprint of only 2.483 MB. The platform integrates a multimodal sensor suite, RGB cameras, Time-of-Flight sensors, microphones, and ambient sensors, and delivers 10 FPS at a continuous autonomy of ~113 minutes on a 200 mAh battery. These results demonstrate the feasibility of a high-performance, privacy-preserving, and socially acceptable assistive device, and highlight how competitive edge AI solutions increasingly demand tightly integrated, multidisciplinary co-design approaches.

Figures

Figures reproduced from arXiv: 2607.16222 by Andrea Giudici, Aurelio Teliti, Christian Veronesi, Diana Trojaniello, Franco Zappa, Giacomo Gervasoni, Mario Cali\`o, Pietro Bartoli.

Figure 2
Figure 2. Figure 2: Render of the ARGO smart eyewear platform. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Render of the smart eyewear. Part I comprises [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Power scheme of the whole system. The core of the power supply is provided by the PMIC. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: High-level representation of the system’s Finite [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Representative samples from the WOTR dataset. Each panel illustrates typical urban scenes containing [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: mAP50–95 versus input size on the WOTR test set for YOLOv11n/s/m in FP32 and INT8 PTQ. INT8 yields consistently lower mAP50–95 across all model families and input resolutions. volatile memory pressure, while maintaining competitive accuracy. Furthermore, with this input size, the network latency is significantly reduced, achieving a 28% reduc￾tion compared to the 320 × 320 configuration, guarantee￾ing the … view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative on-device detection examples pro [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Experimental measurement setup. Two syn [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Power consumption breakdown of the main circuit blocks during one application cycle. Red vertical lines [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Full battery discharge profile obtained with a [PITH_FULL_IMAGE:figures/full_fig_p015_11.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 · 2 linked inside Pith

  1. [1]

    Edge intelligence: Paving the last mile of artificial intelligence with edge comput- ing.Proceedings of the IEEE, 107(8):1738–1762, 2019

    Zhi Zhou, Xu Chen, En Li, Liekang Zeng, Ke Luo, and Junshan Zhang. Edge intelligence: Paving the last mile of artificial intelligence with edge comput- ing.Proceedings of the IEEE, 107(8):1738–1762, 2019

    2019

  2. [2]

    Vivienne Sze, Yu-Hsin Chen, Tien-Ju Yang, and Joel S. Emer. Efficient processing of deep neural networks: A tutorial and survey.Proceedings of the IEEE, 105(12):2295–2329, 2017

    2017

  3. [3]

    A tutorial on human activity recognition using body- worn inertial sensors.ACM Computing Surveys (CSUR), 46(3):1–33, 2014

    Andreas Bulling, Ulf Blanke, and Bernt Schiele. A tutorial on human activity recognition using body- worn inertial sensors.ACM Computing Surveys (CSUR), 46(3):1–33, 2014

    2014

  4. [4]

    Wearable devices for remote monitoring of heart rate and heart rate variability—what we know and what is coming.Sensors, 22(22):8903, 2022

    Navya Alugubelli, Hussam Abuissa, and Attila Roka. Wearable devices for remote monitoring of heart rate and heart rate variability—what we know and what is coming.Sensors, 22(22):8903, 2022

    2022

  5. [5]

    Wearable continuous glucose monitoring sen- sors: a revolution in diabetes treatment.Electronics, 6(3):65, 2017

    Giacomo Cappon, Giada Acciaroli, Martina Vet- toretti, Andrea Facchinetti, and Giovanni Spara- cino. Wearable continuous glucose monitoring sen- sors: a revolution in diabetes treatment.Electronics, 6(3):65, 2017

    2017

  6. [6]

    Alessandra Angelucci and Andrea Aliverti. An imu- based wearable system for respiratory rate estima- tion in static and dynamic conditions: Angelucci and aliverti.Cardiovascular Engineering and Tech- nology, 14(3):351–363, 2023

    2023

  7. [7]

    Multisensor smart eyewear for biomarkers acquisition.IEEE Sensors Journal, 25(21):40565– 40574, 2025

    Ilaria Crupi, Alice Scandelli, Andrea Giudici, Gia- como Gervasoni, Diana Trojaniello, and Federica Villa. Multisensor smart eyewear for biomarkers acquisition.IEEE Sensors Journal, 25(21):40565– 40574, 2025

    2025

  8. [8]

    Open-set recognition of human activities from head- mounted inertial sensor.Sensors, 26(3), 2026

    Angela Cortese, Sarah Solbiati, Alice Scandelli, An- drea Giudici, Niccolò Antonello, Diana Trojaniello, Giacomo Boracchi, and Enrico Gianluca Caiani. Open-set recognition of human activities from head- mounted inertial sensor.Sensors, 26(3), 2026

    2026

  9. [9]

    Aug- mented reality smart glasses: An investigation of technology acceptance drivers.International Jour- nal of Technology Marketing, 11(2):123–148, 2016

    Philipp A Rauschnabel and Young K Ro. Aug- mented reality smart glasses: An investigation of technology acceptance drivers.International Jour- nal of Technology Marketing, 11(2):123–148, 2016

    2016

  10. [10]

    Regazzoni, and Matthias Rauterberg

    Alejandro Betancourt, Pietro Morerio, Carlo S. Regazzoni, and Matthias Rauterberg. The evolu- tion of first person vision methods: A survey.IEEE Transactions on Circuits and Systems for Video Tech- nology, 25(5):744–760, 2015

    2015

  11. [11]

    Your eyes under pres- sure: Real-time estimation of cognitive load with smooth pursuit tracking.Big Data and Cognitive Computing, 9(11):288, 2025

    Pierluigi Dell’Acqua, Marco Garofalo, Francesco La Rosa, and Massimo Villari. Your eyes under pres- sure: Real-time estimation of cognitive load with smooth pursuit tracking.Big Data and Cognitive Computing, 9(11):288, 2025

    2025

  12. [12]

    Egoevgesture: Gesture recognition based on egocentric event camera.arXiv preprint arXiv:2503.12419, 2025

    Luming Wang, Hao Shi, Xiaoting Yin, Kailun Yang, Kaiwei Wang, and Jian Bai. Egoevgesture: Gesture recognition based on egocentric event camera.arXiv preprint arXiv:2503.12419, 2025

    Pith/arXiv arXiv 2025

  13. [13]

    Crafa, Tommaso Polonelli, Marco Carminati, and Michele Magno

    Daniele M. Crafa, Tommaso Polonelli, Marco Carminati, and Michele Magno. Cupido: An analog ultra-low-power and contactless eye blink detector for smart glasses. In2025 IEEE Sensors Applica- tions Symposium (SAS), pages 1–6, 2025

    2025

  14. [14]

    Pettenella, D

    A. Pettenella, D. M. Crafa, J. Spagnoli, L. Merigol, T. Ongarello, and M. Carminati. Development of a compact neural network for ir background esti- mation in wearable eye trackers. In2025 IEEE In- ternational Conference on Metrology for eXtended Reality, Artificial Intelligence and Neural Engineer- ing (MetroXRAINE), pages 1042–1047, 2025

    2025

  15. [15]

    Real-time lighting estimation for aug- mented reality via differentiable screen-space ren- dering.IEEE Transactions on Visualization and Computer Graphics, 29(4):2132–2145, 2022

    Celong Liu, Lingyu Wang, Zhong Li, Shuxue Quan, and Yi Xu. Real-time lighting estimation for aug- mented reality via differentiable screen-space ren- dering.IEEE Transactions on Visualization and Computer Graphics, 29(4):2132–2145, 2022

    2022

  16. [16]

    Optimization methods, challenges, and opportunities for edge inference: A comprehensive survey.Electronics, 14(7), 2025

    Runhua Zhang, Hongxu Jiang, Wei Wang, and Jin- hao Liu. Optimization methods, challenges, and opportunities for edge inference: A comprehensive survey.Electronics, 14(7), 2025

    2025

  17. [17]

    16 Empowering edge intelligence: A comprehensive survey on on-device ai models.ACM Computing Surveys, 57(9):1–39, April 2025

    Xubin Wang, Zhiqing Tang, Jianxiong Guo, Tianhui Meng, Chenhao Wang, Tian Wang, and Weijia Jia. 16 Empowering edge intelligence: A comprehensive survey on on-device ai models.ACM Computing Surveys, 57(9):1–39, April 2025

    2025

  18. [18]

    Fu- sion of computer vision and ai in collaborative robotics: A review and future prospects.Applied Sciences, 15(14), 2025

    Yuval Cohen, Amir Biton, and Shraga Shoval. Fu- sion of computer vision and ai in collaborative robotics: A review and future prospects.Applied Sciences, 15(14), 2025

    2025

  19. [19]

    Intelligent autonomous mobile robot with integrated visual and sensor-based navigation

    Samita Bhandari, Darsh Jobanputra, Aarya Mourya, and Dev Raval. Intelligent autonomous mobile robot with integrated visual and sensor-based navigation. International Research Journal on Advanced En- gineering Hub (IRJAEH), 3(04):1617–1624, April 2025

    2025

  20. [20]

    You only look once: Unified, real- time object detection

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real- time object detection. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 779–788, 2016

    2016

  21. [21]

    Project aria: A new tool for egocentric multi-modal ai research.arXiv preprint arXiv:2308.13561, 2023

    Jakob Engel, Kiran Somasundaram, Michael Goe- sele, Albert Sun, Alexander Gamino, Andrew Turner, Arjang Talattof, Arnie Yuan, Bilal Souti, Brighid Meredith, et al. Project aria: A new tool for egocentric multi-modal ai research.arXiv preprint arXiv:2308.13561, 2023

    Pith/arXiv arXiv 2023

  22. [22]

    Team open smart glasses: An open ecosystem for wearable com- puting

    Team Open Smart Glasses. Team open smart glasses: An open ecosystem for wearable com- puting. https://teamopensmartglasses.com/,

  23. [23]

    Espressif Systems, 2024

    Espressif Systems.ESP32-S3 Series Datasheet: Wi- Fi + Bluetooth LE SoC with Xtensa LX7 Dual-Core Processor. Espressif Systems, 2024. AI Acceler- ation Support via Xtensa PIE (Processor Interface Extension)

    2024

  24. [24]

    Brilliant labs open-source hardware and software repositories

    Brilliant Labs. Brilliant labs open-source hardware and software repositories. https://github.com/ brilliantlabsAR, 2024. Accessed: 2024-11-21

    2024

  25. [25]

    Halo: Open-source ai-powered augmented reality eyewear

    Brilliant Labs. Halo: Open-source ai-powered augmented reality eyewear. https://brilliant. xyz/products/halo, 2024. Accessed: 2024-11- 21

    2024

  26. [26]

    Helios: An extremely low power event- based gesture recognition for always-on smart eye- wear

    Prarthana Bhattacharyya, Joshua Mitton, Ryan Page, Owen Morgan, Ben Menzies, Gabriel Homewood, Kemi Jacobs, Paolo Baesso, David Trickett, Chris Mair, et al. Helios: An extremely low power event- based gesture recognition for always-on smart eye- wear. InEuropean Conference on Computer Vision, pages 168–184. Springer, 2024

    2024

  27. [27]

    A tinyml model for sidewalk obstacle detection: aiding the blind and visually impaired people.Multimedia Tools and Applications, 84(22):25837–25864, 2025

    Ahmed Boussihmed, Khalid El Makkaoui, Ibrahim Ouahbi, Yassine Maleh, and Abdelaziz Chetouani. A tinyml model for sidewalk obstacle detection: aiding the blind and visually impaired people.Multimedia Tools and Applications, 84(22):25837–25864, 2025

    2025

  28. [28]

    Standalone smart glass system for the blind and visually impaired

    Yi-Ler Poy, Shalini Darmaraju, Choon-Hian Goh, and Ban-Hoe Kwan. Standalone smart glass system for the blind and visually impaired. In2024 IEEE 14th Symposium on Computer Applications & Indus- trial Electronics (ISCAIE), pages 239–244. IEEE, 2024

    2024

  29. [29]

    STMicroelectronics, 2024

    STMicroelectronics.STM32N6 Series: High- performance MCUs with dedicated Neural Process- ing Unit. STMicroelectronics, 2024. Neural-ART Accelerator, 600 GOPS, Arm Cortex-M55

    2024

  30. [30]

    Time-of-flight hand-posture recognition us- ing compact normalized histogram

    Pietro Bartoli, Daniele Saporito, Alice Scandelli, Andrea Giudici, Arianna De Vecchi, and Franco Zappa. Time-of-flight hand-posture recognition us- ing compact normalized histogram. In2024 IEEE Sensors Applications Symposium (SAS), pages 1–6, 2024

    2024

  31. [31]

    Infinitime: Multi-sensor wearable bracelet with human body harvesting.Sustainable Computing: Informatics and Systems, 11:38–49, 2016

    Michele Magno, Davide Brunelli, Lukas Sigrist, Renzo Andri, Lukas Cavigelli, Andres Gomez, and Luca Benini. Infinitime: Multi-sensor wearable bracelet with human body harvesting.Sustainable Computing: Informatics and Systems, 11:38–49, 2016

    2016

  32. [32]

    Advancements in context recognition for edge devices and smart eyewear: Sensors and applications.IEEE Access, 13:57062– 57100, 2025

    Francesca Palermo, Luca Casciano, Lokmane Demagh, Aurelio Teliti, Niccolò Antonello, Gia- como Gervasoni, Hazem Hesham Yousef Shalby, Marco Brando Paracchini, Simone Mentasti, Hao Quan, Riccardo Santambrogio, Cédric Gilbert, Manuel Roveri, Matteo Matteucci, Marco Marcon, and Diana Trojaniello. Advancements in context recognition for edge devices and smart...

    2025

  33. [33]

    Jocher, J

    G. Jocher, J. Qiu, and A. Chaurasia. Ultralytics YOLO (version 11.0.0). Computer software, 2023

    2023

  34. [34]

    Energy charac- terization and optimization of image sensing toward continuous mobile vision

    Robert LiKamWa, Bodhi Priyantha, Matthai Phili- pose, Lin Zhong, and Paramvir Bahl. Energy charac- terization and optimization of image sensing toward continuous mobile vision. InProceeding of the 11th annual international conference on Mobile systems, applications, and services, pages 69–82, 2013

    2013

  35. [35]

    A dataset for the visually impaired walk on the road.Displays, 79:102486, 2023

    Haiying Xia, Cong Yao, Yumei Tan, and Shuxiang Song. A dataset for the visually impaired walk on the road.Displays, 79:102486, 2023

    2023

  36. [36]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. InEuropean confer- ence on computer vision, pages 740–755. Springer, 2014

    2014

  37. [37]

    STMicroelec- tronics, 2025

    STMicroelectronics.ST Neural-ART NPU: Sup- ported Operators and Limitations. STMicroelec- tronics, 2025. STEdgeAI 4.0 Documentation

    2025

  38. [38]

    Benchmarking energy and latency in tinyml: A novel method for resource-constrained ai

    Pietro Bartoli, Christian Veronesi, Andrea Giu- dici, David Siorpaes, Diana Trojaniello, and Franco Zappa. Benchmarking energy and latency in tinyml: A novel method for resource-constrained ai. In2025 International Joint Conference on Neural Networks (IJCNN), pages 1–8, 2025. 17

    2025

  39. [2024]

    Accessed: 2024-11-21

    2024