Pith. sign in

REVIEW 3 major objections 5 minor 30 references

HARNode: A Time-Synchronised, Open-Source, Multi-Device, Wearable System for Ad Hoc Field Studies

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Open-source wearable nodes keep ~1 ms sync and hit 98% accuracy with just seven sensors.

desk verdict A genuinely useful open-source wearable platform, but the 1 ms sync claim and the 98% accuracy are both on shaky evidence; the system deserves review, the evaluation does not. read the letter →

arxiv 2506.03219 v2 pith:QBATWASA submitted 2025-06-03 cs.NI

Philipp Lepold , Tobias Röddiger , Michael Beigl This is my paper
classification cs.NI
keywords HumanActivityRecognitionWearableSensorsSensorNodesOpen-SourceHardwareMulti-DeviceSystemsStairWalkingPredictionExoskeletonTimeSynchronisation
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 presents HARNode, a fully open-source wearable sensor system built from off-the-shelf parts that streams data over Wi-Fi and uses NTP (Network Time Protocol) to keep many nodes synchronized to about one millisecond. The central claim is that this removes common barriers to dense field studies: each node costs under €48, setup of eleven nodes takes under five minutes per participant, and a single access point could support roughly a hundred nodes. In a proof-of-concept study with ten participants, a random forest trained on the over-provisioned eleven-node dataset classified walking versus approaching stairs with ~98% accuracy, and the best seven-node subset matched the full set, with even three well-chosen nodes reaching 98.06%. The intended consequence is that researchers can deploy many sensors first and then select informative placements afterward, rather than guessing where to put one or two IMUs.

What carries the argument

The load-bearing mechanism is sensor overprovisioning combined with NTP-aligned streaming. Overprovisioning means wearing more sensors than one expects to need, here eleven nodes in fixed positions, so that all $2^{11}$ subsets can be evaluated by an exhaustive random-forest search. NTP-based time synchronization, with round-trip interpolation and minute-by-minute re-synchronization, is what makes features from different body locations comparable; without consistent timestamps the fused features would be corrupted by inter-node drift. The supporting stack is an ESP32-S3-based node, UDP packets of 30 samples (avoiding fragmentation), and a server that timestamps all streams against the same synchronized clocks.

What would settle it

Measure each HARNode's NTP-reported offset and round-trip time, plus the arrival jitter at the server, during an eleven-node recording; if inter-node timestamp disagreement exceeds a few milliseconds, the fusion claim collapses. Separately, retrain the random forest with non-overlapping windows and subject-exclusive splits: if accuracy falls far below 98%, the reported gain from overprovisioning is largely an artifact of information leakage.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a wearable human-activity-recognition system built from cheap, off-the-shelf components can achieve near-laboratory-grade multi-node time alignment in the field. Each node synchronizes to a local NTP server before streaming UDP packets, reaching roughly 1 ms average (5 ms maximum) clock agreement, and re-synchronizes every minute to hold drift at the microsecond level. Using eleven nodes worn by ten participants, the authors train random-forest classifiers for every subset of the eleven positions (2,048 models) to detect the transition from walking to approaching stairs. The best seven-node combination reaches 98.38% accuracy, essentially matching the 98.01% of all eleven nodes; even three nodes (left foot, left wrist, waist) reach 98.06%. The authors take this as evidence that sensor overprovisioning can replace intuition-based placement: deploy a dense array, then search for the smallest informative subset.

Load-bearing premise

The whole multi-node fusion argument assumes the NTP-based time alignment really holds at about one millisecond when many nodes stream over a single Wi-Fi access point; the paper cites an external figure for this and does not measure clock offsets or packet-delay variation in its own setup.

Editorial extensions

If this is right

  • Field researchers can run dense multi-node studies (up to roughly 100 nodes per access point by the paper's throughput estimate) at under €48 per node and under five minutes of donning time per participant.
  • In exoskeleton design, the results suggest that three to seven well-chosen IMU locations are enough for stair-approach detection, so an assistive device need not instrument the whole body.
  • Because the hardware, firmware, and server code are released open source, other groups can reproduce or extend the platform without buying a commercial motion-capture system.
  • The exhaustive subset search provides a template for turning a general HAR dataset into placement recommendations, replacing intuitive sensor-site selection with measured performance.

Reading between the lines

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

  • If the ~1 ms NTP alignment holds under load, the platform could serve as a low-cost testbed for time-critical multi-node fusion such as gait-phase estimation, where even a few milliseconds of skew corrupts cross-node features.
  • The reported ~98% figures likely overstate real-world performance: windows overlap by 75% and class balancing happens before the 70/30 split, so adjacent training and test windows share information; the authors' own leave-one-subject-out accuracy (~91%) is a more conservative estimate.
  • The system's I2C port and modular PCB invite adding other sensors, which might reveal further placement effects beyond the IMU-only study.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces HARNode, an open-source wearable multi-node sensing platform built from an M5Stack AtomS3 (ESP32-S3), a Bosch BMX160 9-axis IMU, a BMP388 pressure/temperature sensor, a custom PCB, and a 3D-printed housing with Velcro straps. Data are streamed over Wi-Fi as UDP packets to a NodeJS server, and the authors claim NTP-based synchronization to roughly 1 ms across nodes. To demonstrate the system, they report a proof-of-concept study with 10 participants wearing 11 HARNodes each, classifying walking versus approaching stairs with a random forest; they report approximately 98% accuracy for both 7 selected nodes and the full 11-node set, and interpret this as evidence for sensor overprovisioning and for HARNode's utility in field studies.

Significance. If the synchronization and evaluation claims hold, HARNode would be a valuable open-source contribution: it is low-cost (~48 EUR per node), fast to deploy (under 5 minutes for 11 nodes in the reported study), and designed for dense on-body placement, addressing a real gap in accessible field-deployable wearable sensing. The open-source release of hardware, firmware, and server software is a concrete strength, as is the systematic comparison over all 2^11 sensor combinations. However, the two load-bearing quantitative claims are currently not adequately supported: the 1 ms synchronization figure is borrowed from the NTP literature rather than measured on this system, and the 98% accuracy figures are selection results obtained through an evaluation procedure with information leakage. These issues materially affect the paper's central contribution and require correction before the claims can be accepted.

major comments (3)
  1. [§3.3.1] The 'roughly 1 ms' synchronization claim is asserted, not demonstrated. The text cites Mills [20] for '1 ms in average (5 ms max.)', but [20] characterizes general Internet NTP behavior and is not a measurement of this firmware on ESP32-S3 nodes over one 2.4 GHz access point while all nodes stream UDP packets at the §5.1 rate. No clock offsets, round-trip times, jitter, or packet-delay variation are reported for HARNodes under field load. Because §5.2 explicitly relies on these timestamps to align and interpolate all 11 sensor streams, the synchronization accuracy is load-bearing for the multi-node fusion results. The authors should add a dedicated synchronization measurement, for example a simultaneous physical trigger captured by every node or a comparison of node-reported timestamps against server-side arrival times, and report the resulting offset distribution and jitter.
  2. [§5.2–§6] The headline '7 nodes ≈ 98%' and '3–11 nodes give nearly the same accuracy' results are fitted selection outcomes, not predictions. The best subset for each sensor count was chosen by maximizing accuracy over all 2^11 = 2048 models on the same dataset, so the values in Table 3 and Figure 5 are optimistically biased by construction; the LOOCV means (90.99%, 90.53%, 88.74%) are reported only after selection and are not valid estimates for the chosen subsets either. A nested cross-validation or a fully held-out participant set is needed to estimate what accuracy a researcher should expect from the sensor-selection procedure, and the abstract and conclusion should describe the 98% figures as retrospective screening results unless that experiment is added.
  3. [§5.2] The evaluation leaks information between training and test windows. The continuous streams are segmented into windows of 25 samples with 75% overlap, and the 70/30 train/test split is applied after segmentation, so overlapping windows from the same recorded session can appear in both training and test sets. Balancing the class distribution before the split further increases the leakage. The classification should be evaluated with a participant-level or session-level split applied before windowing, or with group-wise cross-validation that keeps all windows from one recording together, and the reported confidence intervals should reflect that grouping.
minor comments (5)
  1. [§3.3.1] The text conflates 'about 1 ms average', '5 ms max', and 'microsecond level' RTC deviation; these are different quantities (offset, jitter, drift) and should be defined and measured separately.
  2. [§5.1] The sentence '10 right-footed, 1 left-footed' is inconsistent with the stated ten participants; please correct the numbers.
  3. [Figure 5] The y-axis label is not fully visible and there are no error bars or per-subject spread, making it impossible to judge whether the accuracy differences across node counts are meaningful.
  4. [Table 3] The LOOCV means are only given in prose; they should appear in the table alongside the in-sample accuracies, with a clarifying note on why they are about 7 points lower.
  5. [Throughout] There are several language and typographical issues, including 'under-represent class', 'wearablity', and 'propr' in Table 1; a proofread would improve readability.

Circularity Check

1 steps flagged · score 6.0 of 10

The overprovisioning claim that seven nodes match eleven is derived by taking the maximum accuracy over 2048 node-combination models on the same test split, so the headline 98% figure is forced by the selection procedure rather than independently validated.

  1. fitted input called prediction [Section 6 (Evaluation), first paragraph and the paragraph following Table 3.]
    "we employed sensor overprovisioning and trained random forest models for binary classification ("walking" and "towards stairs") across all possible combinations of sensor positions (2^11 = 2048 models). For every amount of sensor (1-11), the model with the highest accuracy was calculated... The best model with all 11 sensor nodes achieves 98.01% accuracy ... while the highest overall accuracy of 98.38% ... is achieved with only 7 nodes."

    The reported 'seven nodes ≈ 98%, matching all 11' result is the upper envelope of the search: for each sensor count, the subset with the highest accuracy on the test split is selected, and that same selected accuracy is then reported as the finding. The claim is therefore entailed by the selection procedure, not measured as a representative or predictive property of seven-node configurations. Choosing the maximum over 2048 combinations and then concluding that a small subset 'performs comparably' to all eleven presents the selection criterion itself as evidence for sensor overprovisioning.

full rationale

No genuine derivation chain in this systems paper reduces to its own inputs. The hardware, firmware, and open-source claims are self-contained and externally checkable. The NTP synchronization claim ('about 1 ms in average (5 ms max.) [20]') is a correctness/evidence gap, not circularity: it imports an external specification from Mills' general NTP paper without measuring this system's nodes, but it is not defined in terms of the conclusion. The only construction-level circularity is in the validation study: the paper selects the accuracy-maximizing node subset on the test split and then presents the resulting maximum as demonstrating that 3 to 11 nodes 'achieve nearly the same accuracy.' That is a fitted result called a finding, so the overprovisioning conclusion is forced by the selection procedure rather than independently supported. Additional methodological concerns (75% overlapping windows and class balancing before the 70/30 split) further inflate the reported accuracies, but these are data-leakage issues rather than circularity. Overall the core system contribution stands independently; the circular step is confined to the evaluation's headline accuracy claim.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on unverified domain assumptions about NTP accuracy in the HARNode network and on ML design choices that are not independently validated. No new physical entities are introduced.

free parameters (4)
  • Sensor subset used in overprovisioning comparison = 3 nodes (left foot, left wrist, waist) 98.06%; 7 nodes 98.38%
    The best combination for each sensor count was selected as the maximum over all C(11,n) models evaluated on the same test split, making this a fitted selection rather than a held-out prediction.
  • Segmentation window length and overlap = 25 samples (150 ms), 75% overlap
    Chosen without reported sensitivity analysis; windowing determines feature counts and induces overlap between train and test windows under a random split.
  • Sampling rate = 166.67 Hz
    Chosen as a compromise between motion detection, battery life, and synchronization; affects all downstream features and accuracy.
  • Random forest hyperparameters = not reported
    No tree count, depth, or split criteria are given, so the reported accuracies cannot be reproduced exactly.
assumptions (5)
  • domain assumption NTP over a local Wi-Fi network yields roughly 1 ms average (5 ms max) synchronization error for the HARNode deployment.
    Invoked in Section 3.3.1 and the abstract; attributed to [20] but never measured for HARNodes.
  • domain assumption Rough placement of the IMU over the intended body segment, without calibration fixtures, yields usable orientation and position for HAR features.
    Stated in Section 4 step 2; no orientation-error analysis is provided.
  • domain assumption Manual video labelling defines correct ground-truth boundaries for 'towards stairs' versus 'walking'.
    Used in Section 5.2; no inter-rater reliability or label uncertainty is reported.
  • domain assumption Class balancing by random sampling of the under-represented class preserves the test distribution.
    Section 5.2; if balancing is done before the 70/30 split, duplicated windows can appear in both train and test sets.
  • domain assumption An effective 2.4 GHz Wi-Fi throughput of at least 5 Mbit/s is available with more than 100 associated nodes.
    Used in Section 3.4 to derive the roughly 103-node limit; not benchmarked in this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HARNode: A Time-Synchronised, Open-Source, Multi-Device, Wearable System for Ad Hoc Field Studies." pith.science (2026). https://pith.science/paper/QBATWASA

@misc{pith2026250603219,
  author       = {Pith},
  title        = {Pith review of: HARNode: A Time-Synchronised, Open-Source, Multi-Device, Wearable System for Ad Hoc Field Studies},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QBATWASA}},
  note         = {Machine review of arXiv:2506.03219}
}
read the original abstract

Human activity recognition (HAR) research often lacks accessible, comprehensive field data. Commercial systems are rarely open source, hard to expand, and limited by issues like node synchronisation, data throughput, unclear sensor placement, complexity, and high cost. As a result, researchers typically use only a few intuitively placed sensors and conduct limited field trials. HARNode overcomes these challenges with a fully open-source hardware and software platform. Each node includes an ESP32-S3 module (AtomS3), a 9-axis IMU (Bosch BMX160), pressure and temperature sensors (Bosch BMP388), a display, and an I2C port. Data is streamed via Wi-Fi, with NTP-based time synchronisation achieving roughly 1 ms accuracy. The system runs for up to 8 hours and is built using off-the-shelf parts, a simple online PCB service, and a compact 3D-printed housing with Velcro straps, enabling flexible and scalable body placement while requiring little hardware knowledge. In a study with ten subjects wearing eleven HARNodes each, setup took under five minutes per person. A random forest classifier distinguished walking from stair-climbing transitions, showing the benefits of sensor-overprovisioning: Seven nodes achieved approx. 98% accuracy, matching the performance of all eleven. These findings confirm HARNode's value as a fast-deploying, scalable tool for field-based HAR research and optimised sensor placement.

Figures

Figures reproduced from arXiv: 2506.03219 by the authors.

Figure 1
Figure 1. (A) Schematic diagram of the sensor node, showing the sensors, and power circuitry; (B) Assembled sensor node PCB [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. (A) Rendering of the 3D-printed HARNode housing; [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Assembly process of HARNode: (A) The PCB is plugged into the 3D-printed housing; (B) the battery [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Sensor node placement on participant. For our vali [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 6
Figure 6. Figure 6: Confusion matrix for binary classification ( [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 22 canonical work pages

  1. [20]

    David L Mills. 2002. Internet time synchronization: the network time protocol. IEEE Transactions on communications 39, 10 (2002), 1482–1493

    work page 2002

  2. [1]

    2009. IEEE Standard for Information technology– Local and metropolitan area networks– Specific requirements– Part 11: Wireless LAN Medium Access Control (MAC)and Physical Layer (PHY) Specifications Amendment 5: Enhancements for Higher Throughput. IEEE Std 802.11n-2009 (Amendment to IEEE Std 802.11-2007 as amended by IEEE Std 802.11k-2008, IEEE Std 802.11r...

    arXiv 2009

  3. [2]

    Movisens GmbH

    2025. Movisens GmbH. https://www.movisens.com/. Accessed: May 21, 2025

    work page 2025

  4. [3]

    MVN Awinda

    2025. MVN Awinda. https://www.movella.com/products/motion-capture/xsens- mvn-awinda. Accessed: May 21, 2025

    work page 2025

  5. [4]

    Noraxon: Innovative Biomechanics Tools for Research & Clinical Use

    2025. Noraxon: Innovative Biomechanics Tools for Research & Clinical Use. https://www.noraxon.com/. Accessed: May 21, 2025

    work page 2025

  6. [5]

    The Opal: a research-grade wearable sensor built for ultimate control

    2025. The Opal: a research-grade wearable sensor built for ultimate control. https://biofittechs.com/apdm-wearable. Accessed: May 21, 2025

    work page 2025

  7. [6]

    Wearable and Real-time Kinematics Estimates with OpenSense

    2025. Wearable and Real-time Kinematics Estimates with OpenSense. https://opensimconfluence.atlassian.net/wiki/spaces/OpenSim/pages/53084280/ Wearable+and+Real-time+Kinematics+Estimates+with+OpenSense. Accessed: May 21, 2025

    work page arXiv 2025

  8. [7]

    Laith Alzubaidi, Jinshuai Bai, Aiman Al-Sabaawi, Jose Santamaría, Ahmed Shihab Albahri, Bashar Sami Nayyef Al-Dabbagh, Mohammed A Fadhel, Mohamed Manoufali, Jinglan Zhang, Ali H Al-Timemy, et al . 2023. A survey on deep learning tools dealing with data scarcity: definitions, challenges, solutions, tips, and applications. Journal of Big Data 10, 1 (2023), 46

    work page 2023

Show all 30 references
  1. [8]

    Apple Inc. 2024. How to set up a Personal Hotspot on your iPhone or iPad . https: //support.apple.com/en-us/111785 Accessed: 2025-05-21

    2024

  2. [9]

    Giorgio Biagetti, Michele Sulis, Laura Falaschetti, and Paolo Crippa. 2025. High- Accuracy Clock Synchronization in Low-Power Wireless sEMG Sensors. Sensors 25, 3 (2025), 756. doi:10.3390/s25030756

    2025 doi

  3. [10]

    Chiara Contoli, Valerio Freschi, and Emanuele Lattanzi. 2024. Energy-aware hu- man activity recognition for wearable devices: A comprehensive review.Pervasive and Mobile Computing (2024), 101976

    2024

  4. [11]

    Mardini, Parisa Rashidi, and Todd Manini

    Anis Davoudi, Mamoun T. Mardini, Parisa Rashidi, and Todd Manini. 2021. The Effect of Sensor Placement and Number on Physical Activity Recognition and Energy Expenditure Estimation in Older Adults. JMIR mHealth and uHealth 9, 5 (2021), e23681. doi:10.2196/23681

    2021 doi

  5. [12]

    Lars Eggert, Godred Fairhurst, and Greg Shepherd. 2017. UDP Usage Guidelines. https://datatracker.ietf.org/doc/html/rfc8085. RFC 8085

    2017

  6. [13]

    Espressif Systems. 2021. ESP32-S3 Series Datasheet. https://www.espressif.com/ sites/default/files/documentation/esp32-s3_datasheet_en.pdf Accessed: 2025-05- 26

    2021

  7. [14]

    Manuel Gil-Martín, Javier López-Iniesta, Fernando Fernández-Martínez, and Rubén San-Segundo. 2023. Reducing the impact of sensor orientation variability in human activity recognition using a consistent reference system. Sensors 23, 13 (2023), 5845

    2023

  8. [15]

    Zhiqing Hong, Zelong Li, Shuxin Zhong, Wenjun Lyu, Haotian Wang, Yi Ding, Tian He, and Desheng Zhang. 2024. Crosshar: Generalizing cross-dataset human activity recognition via hierarchical self-supervised pretraining. Proceedings of the ACM on Interactive, Mobile, Wearable and...

    2024

  9. [16]

    Zawar Hussain, Michael Sheng, and Wei Emma Zhang. 2019. Different approaches for human activity recognition: A survey. arXiv preprint arXiv:1906.05074 (2019)

    2019 arXiv

  10. [17]

    Akim Kapsalyamov, Prashant K Jamwal, Shahid Hussain, and Mergen H Ghayesh

  11. [18]

    Ahsan Raza Khan, Habib Ullah Manzoor, and Ahmed Zoha. 2022. A Privacy and Energy-Aware Federated Framework for Human Activity Recognition. Sensors 23, 23 (2022), 9339. doi:10.3390/s23239339

    2022 doi

  12. [19]

    Kai Kunze and Paul Lukowicz. 2014. Sensor placement variations in wearable activity recognition. IEEE Pervasive Computing 13, 4 (2014), 32–41

    2014

  13. [21]

    Luís Moreira, Joana Figueiredo, João Cerqueira, and Cristina P Santos. 2022. A review on locomotion mode recognition and prediction when using active orthoses and exoskeletons. Sensors 22, 19 (2022), 7109

    2022

  14. [22]

    Haque, Yan Yan, and Anne H.˜H

    Jianyuan Ni, Hao Tang, Syed T. Haque, Yan Yan, and Anne H.˜H. Ngu. 2024. A Survey on Multimodal Wearable Sensor-Based Human Action Recognition.arXiv e-prints (2024). arXiv:2404.15349

    2024 arXiv

  15. [23]

    Mattia Pesenti, Giovanni Invernizzi, Julie Mazzella, Marco Bocciolone, Alessandra Pedrocchi, and Marta Gandolla. 2023. IMU-based human activity recognition and payload classification for low-back exoskeletons. Scientific Reports 13, 1 (2023), 1184

    2023

  16. [24]

    Patrick Slade, Ayman Habib, Jennifer L Hicks, and Scott L Delp. 2021. An open- source and wearable system for measuring 3D human motion in real-time. IEEE Transactions on Biomedical Engineering 69, 2 (2021), 678–688

    2021

  17. [25]

    Matteo Sposito, Tommaso Poliero, Christian Di Natali, Marianna Semprini, Giac- into Barresi, Matteo Laffranchi, Darwin Gordon Caldwell, Lorenzo De Michieli, and Jesús Ortiz. 2022. Exoskeletons in elderly healthcare. In Internet of Things for Human-Centered Design: Application ...

    2022

  18. [26]

    Akihisa Tsukamoto, Naoto Yoshida, Tomoko Yonezawa, Kenji Mase, and Yu Enok- ibori. 2023. Where Are the Best Positions of IMU Sensors for HAR?-Approach by a Garment Device with Fine-Grained Grid IMUs. In Adjunct Proceedings of the 2023 ACM International Joint Conference on Perv...

    2023

  19. [27]

    https://www.bioxgroup.dk/products-biox-armband/. [n. d.]. BioXBands. Ac- cessed: May 21, 2025

    2025

  20. [28]

    Clint Zeagler. 2017. Where to wear it: functional, technical, and social consid- erations in on-body location for wearable technology 20 years of designing for wearability. In Proceedings of the 2017 ACM International Symposium on Wearable Computers. 150–157

    2017

  21. [29]

    Shibo Zhang, Yaxuan Li, Shen Zhang, Farzad Shahabi, Stephen Xia, Yu Deng, and Nabil Alshurafa. 2022. Deep learning in human activity recognition with wearable sensors: A review on advances. Sensors 22, 4 (2022), 1476

    2022

  22. [2019]

    IEEE Access 7 (2019), 95075–95086

    State of the art lower limb robotic exoskeletons for elderly assistance. IEEE Access 7 (2019), 95075–95086

    2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.