RARR : Robust Real-World Activity Recognition with Vibration by Scavenging Near-Surface Audio Online

Alyssa Weakley; Blake Brown; Dong Yoon Lee; Hui Wei; Keyana Carrion; Shijia Pan

REVIEW 4 major objections 6 minor 20 references

RARR : Robust Real-World Activity Recognition with Vibration by Scavenging Near-Surface Audio Online

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that publicly available ASMR recordings can substitute for large labeled vibration datasets in privacy-preserving home activity recognition.

desk verdict Novel ASMR-pretraining idea for vibration HAR, but the headline 13% improvement is not reproducible from the paper's own numbers. read the letter →

arxiv 2508.21167 v1 pith:5MSGODEP submitted 2025-08-28 cs.SD cs.LG

Dong Yoon Lee , Alyssa Weakley , Hui Wei , Blake Brown , Keyana Carrion , Shijia Pan This is my paper

classification cs.SDcs.LG

keywords activityrecognitionstructuralvibrationsensingASMRaudiopretrainingtransferlearningvariationalautoencodermultitasksmarthomemonitoringaginginplace

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 claims that publicly available ASMR recordings can substitute for large labeled vibration datasets when building activity recognizers for privacy-preserving home monitoring. The physical motivation is that a footfall creates both airborne sound and floor-borne vibration, and ASMR videos captured with boundary microphones emphasize near-surface sounds that resemble what a floor sensor feels. The authors pretrain a variational autoencoder on keyword-curated ASMR audio of daily activities, then fine-tune only the activity-recognition branch on a small amount of real vibration data from one participant. On a four-participant home-suite deployment, the resulting system outperforms three baselines for every participant, with a reported average improvement of about 13%. If true, this would let unobtrusive vibration-based monitoring be deployed to new homes without the annotation burden that currently blocks it.

What carries the argument

The load-bearing object is the multitask variational autoencoder with separated mean and variance latents, combined with a task-selective fine-tuning policy and a low-rank 1D CNN latent adapter (the ASMR-Vibration Adapter). The mean latent is the only input to the activity classifier, so fine-tuning the classifier shifts the modality-specific representation from near-surface audio to on-surface vibration; the variance latent, trained by the reconstruction task, stays frozen and keeps the model tolerant of different users, environments, and sensor conditions. The ASMR dataset itself is the second piece of machinery: keyword queries turn publicly available recordings into a weakly labeled near

What would settle it

Train the same RARR pipeline on ASMR clips whose activity labels have been shuffled. If the shuffled-label model matches RARR's accuracy on unseen participants, the activity-specific audio content is not what transfers and the pretraining claim collapses. Separately, deploy the system on different flooring types (carpet vs. hardwood) and if accuracy on new users drops to the simple no-pretraining baseline, the claimed robustness across environments fails.

Watch

Extended reading notes

Core claim

The central claim is that acoustic and vibration sensing of the same activity are close enough in structure that a model pretrained on the former can be transferred to the latter with very little labeled vibration data. RARR does this with a multitask variational autoencoder that splits its latent code: the mean latent carries categorical activity information and feeds a temporal convolutional classifier, while the variance latent, used for signal reconstruction, absorbs user- and environment-specific variation. Pretraining uses 3,066 overlapping 30-second clips of ASMR audio collected by searching for activity keywords; fine-tuning uses vibration data from a single participant's five activi

Load-bearing premise

The load-bearing premise is that keyword-matched ASMR recordings, made by unknown people with unknown microphones, sound enough like what a floor sensor feels that the same activity leaves a shared learnable pattern in both signals—and that the search-engine labels are trustworthy without manual checks.

Editorial extensions

If this is right

Vibration-based activity monitoring could be installed in a new home with labeled data from a single user or a few short sessions, rather than weeks of annotated floor signals.
Public ASMR libraries become a reusable pretraining resource for privacy-preserving sensing, reducing the need for large shared vibration datasets.
Task-selective fine-tuning preserves the variance robustness learned during pretraining, which should make accuracy more stable across unseen users than full fine-tuning.
The framework can extend to other activities as long as keyword-matched near-surface audio is available online, so the activity set is not fixed by the sensor's physics.
If the accuracy holds at community scale, remote caregivers could receive reliable alerts about walking, showering, and medication events without cameras or microphones in the home.

Reading between the lines

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

Editorial inference: the same modality-proximity trick—pretraining on abundant online audio, fine-tuning on scarce structural vibration—could be tested for other indirect sensors, such as accelerometers on appliances or wearables, where the physical coupling to sound is weaker.
Editorial inference: the keyword-labeling step is an uncontrolled variable; filtering ASMR clips by spectral cues characteristic of the target activity could reduce label noise and likely push accuracy beyond the reported 13%.
Editorial inference: the mean/variance latent split suggests a general recipe for domain adaptation in sensing—only the discriminative stream is tuned while the generative stream is frozen—and that recipe could be evaluated on unrelated modality pairs to see where it stops working.

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.

Desk Editor's Note

read the letter

Quick take: the core idea—pretraining a multitask VAE on public ASMR audio and task-selectively fine-tuning on a tiny vibration dataset—is genuinely new and worth discussing. The paper is clearly written, the design choices (freezing most layers, adding a low-rank adapter) are plausible, and the authors are honest about the proof-of-concept scope in the discussion.

But the quantitative evidence does not hold up. Section 3.3 reports per-participant accuracies of 71, 75, 67, 50. Those average to 65.75, not 72. The claimed standard error of 18% is also off by a factor of ~4. And the repeated claim of 'around 13% average improvement' over baselines is tied to no baseline definition. Using the paper's own Fig. 4, the best baseline (A2V-VAE) averages 37.25%, so the improvement would be roughly 28.5 points (or ~76% relative), neither of which is 13. This is an internal inconsistency, not just a weak external validation.

The evaluation is also underpowered: four participants, one training subject, and no released code or data. The ASMR dataset is assembled by keyword search with no manual label validation. Participant 4 gets 0% on all three baselines, which suggests a sensor or label issue that goes undiscussed. That said, the transfer-learning framing is sound, and the circularity concern doesn't apply—test participants are held out and pretraining doesn't use vibration labels.

Bottom line: the idea deserves a serious look, but the claims as written are not supported by the data. I'd send this to peer review with a strong expectation of major revision—particularly fixing the arithmetic, defining the baseline for the 13% figure, and releasing the data/code. If those numbers don't resolve, the contribution reduces to a plausible recipe with an unverified payoff.

For a reading group: maybe, if the group is interested in cross-modal pretraining for sensing. I wouldn't cite it in its current form.

Referee Report

4 major / 6 minor

Summary. The paper proposes RARR, a framework for vibration-based activity recognition that pretrains a multitask variational autoencoder (VAE) on publicly available ASMR audio retrieved by keyword search, then fine-tunes only the activity-recognition branch (with a low-rank latent adapter) on a small amount of structural-vibration data from one participant, and evaluates on three held-out participants. The architecture combines a VAE encoder with a temporal convolutional network (TCN) for activity classification and a decoder for STFT reconstruction; the claimed contribution is a 13% average accuracy improvement over three baselines, with an overall mean accuracy of 72% (Section 1, Section 3.3, Conclusion). The paper also discusses the potential of the approach for remote monitoring of older adults living alone.

Significance. The direction is timely and creative: using freely available ASMR recordings as a proxy pretraining source for structural-vibration sensing could substantially reduce labeled-data requirements for in-home activity recognition. The multitask VAE plus task-selective fine-tuning idea is sensible, and the paper correctly identifies that indirect sensing signals are hard for humans to annotate, motivating transfer from more abundant audio data. I found no clear fitting-vs-prediction circularity: the reported test accuracies are on held-out participants, and the ASMR pretraining objective does not use vibration labels. However, the quantitative evidence as reported is internally inconsistent, and the evaluation is very small (four participants, one training participant, no released code or data). The significance of the claimed result is therefore not established in the current manuscript.

major comments (4)

[Section 1, Section 3.3, Figure 4] The headline numbers do not reproduce. The four RARR per-participant accuracies (71%, 75%, 67%, 50%) average to 65.75%, not the reported 72%; the standard error of these four values is about 5.5 percentage points, not 18%. The best baseline, A2V-VAE, averages (61+50+38+0)/4 = 37.25%. The absolute improvement is therefore 28.5 percentage points and the relative improvement is about 76%, not 13%. Against Simple-VAE (mean 36%) or Pretrained-VAE (mean 34%), the relative improvements are about 83% and 93%, respectively. No definition of 'average improvement' is given. Because the '13% improvement' and '72% accuracy' appear in the introduction and conclusion as the central quantitative claims, this internal inconsistency is load-bearing and must be corrected.
[Section 3.3, Participant 4] Participant 4 has 0% accuracy for all three baselines and only 50% for RARR. This extreme outlier strongly influences the aggregated comparison: without Participant 4, A2V-VAE's mean accuracy rises from 37.25% to about 49.7%, and RARR's mean rises from 65.75% to 71%. The paper does not discuss whether this indicates a sensor malfunction, missing classes, label misalignment, or a qualitatively different signal distribution. The authors should report per-activity/per-class results, analyze Participant 4 separately, and show aggregate results both with and without this participant.
[Section 3.1, Section 3.2] The experimental protocol is under-specified and contradictory. Section 3.1 says participants 'conducted 30 different sequences involving 4 (of possible 23; including walking, showering, medication taking, and medication refilling) activities,' while Section 3.2 says fine-tuning uses 'all five activity sequences from participant 1.' The relationship between '30 sequences,' 'four activities,' and 'five activity sequences' is unclear. The paper also does not state how many test samples per participant were used to compute the accuracies in Figure 4, which is essential for interpreting the reported standard error. Please clarify the protocol and report test-set sizes.
[Section 2.1, Section 3.1] The central premise is that ASMR recordings retrieved by keyword search provide activity-aligned audio suitable for pretraining a vibration-based HAR model. However, the paper provides no validation of the ASMR labels: no manual verification, no count of videos/creators per label, no information about microphone types or recording environments, and no spectrogram or embedding-level evidence that the ASMR recordings are acoustically similar to the target floor-borne vibrations. Since the entire transfer-learning claim rests on this premise, please provide such evidence or an ablation that measures the contribution of ASMR pretraining quality (e.g., manual label audit, number of creators, or a comparison against mismatched audio pretraining).

minor comments (6)

[Section 3.1] The text says 'three common and important home activities' but then lists four: walking, showering, medication intake, and medication refilling.
[Section 2.2 / Figure 3] The figure labels the second task as 'Recreation' while the text consistently uses 'STFT signal recreation' or 'reconstruction.' Please unify terminology (likely 'reconstruction').
[Section 3.1] The ASMR dataset description says clips are 30 seconds with 15-second overlap, but the 60/40 train/validation split is not described at the video/creator level. If overlapping clips from the same video appear in both training and validation, this could cause leakage. Please specify the split strategy.
[Section 2.1] The claim that ASMR is captured by boundary microphones is supported only by a non-peer-reviewed web source [1]. Please qualify this or provide a technical reference.
[Abstract] Grammar: 'One in four people dementia live alone' should read 'One in four people with dementia live alone.' Similar phrasing appears in the introduction.
[Section 5.1] The discussion of masked and conditional autoencoders is generic and does not connect to the experiments or to RARR's design. Consider shortening or making the connection explicit.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: RARR's central claim is an empirical held-out benchmark comparison; internal 13%/72% arithmetic inconsistency is a correctness issue, not circularity.

full rationale

The paper's derivation chain is: (1) curate ASMR audio by keyword search; (2) pretrain a multitask VAE on STFTs of that audio with reconstruction and activity-prediction objectives (labels from search keywords); (3) fine-tune only the TCN final layer and add a low-rank 1D CNN adapter using participant 1's labeled vibration data; (4) evaluate activity accuracy on held-out participants 2-4. None of these steps defines the test quantity in terms of the fitted quantity. The ASMR labels are noisy, but the reported accuracies are computed against independently established ground truth from the Home Health Simulation Suite and on unseen participants, so the claimed 13% improvement is an experimental result, not a consequence of the model's construction. The paper's self-citations (e.g., [5], [6], [15], [17]) are used as background or physics context and are not load-bearing for the main result; no uniqueness theorem or ansatz is imported from them. The manuscript does contain a stated limitation in Section 5.2 (larger-scale deployments remain untested) and an internal arithmetic inconsistency in Section 3.3: the listed RARR per-participant accuracies (71%, 75%, 67%, 50%) average to 65.75%, not 72%, and the reported 18% standard error does not match those values; the 13% improvement figure in the introduction/conclusion is likewise not reproduced from Figure 4 by any standard metric. However, these are reproducibility/correctness concerns about empirical reporting, not circularity: no equation or fitted parameter is renamed as a prediction, and the evaluation uses held-out participants. Therefore the circularity score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 1 invented entities

The central claim depends on the transferability of ASMR audio representations to floor vibration, on unvalidated keyword-derived labels, and on the representational assumptions of the VAE. These are domain assumptions rather than fitted parameters; the exact hyperparameters and STFT settings are not disclosed.

free parameters (2)

VAE latent dimensions and adapter ranks = unreported
The low-rank 1D CNN adapter and the VAE latent dimensionality are chosen by hand; values are not disclosed, affecting reproducibility.
STFT window and hop sizes = unreported
The STFT parameters of the input are not stated; they determine the input representation and are free choices.

assumptions (4)

domain assumption ASMR audio from online videos approximates near-surface acoustic signals of the target home activities.
Section 2.1 assumes boundary microphones in ASMR recordings capture the same pressure-zone physics as the target surface, but this is not validated against the actual corpus.
domain assumption Keyword-based querying of videos yields reliable activity labels for the ASMR dataset.
Section 2.1 and 3.1: labels are derived from search terms without manual verification, annotation, or quality control.
domain assumption Activity-induced vibration signals share latent structure with near-surface audio such that VAE pretraining transfers.
Section 2.2: the entire transfer relies on this coupling; no direct analysis is provided beyond the accuracy comparison.
standard math The variational autoencoder reparameterization and training objective are standard and correct.
Section 2.2 cites [3] (Doersch, VAE tutorial) for the VAE formulation.

invented entities (1)

ASMR-Vibration Adapter (AVA)
purpose: Low-rank 1D CNN adapter to shift the latent representation from near-surface audio to on-surface vibration during fine-tuning.
The adapter is part of the proposed architecture; its benefit is only demonstrated in the in-house experiment, with no external validation or ablations.

how reviews work

0 comments

Cite this review

Pith. "Pith review of RARR : Robust Real-World Activity Recognition with Vibration by Scavenging Near-Surface Audio Online." pith.science (2026). https://pith.science/paper/5MSGODEP

@misc{pith2026250821167,
  author       = {Pith},
  title        = {Pith review of: RARR : Robust Real-World Activity Recognition with Vibration by Scavenging Near-Surface Audio Online},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5MSGODEP}},
  note         = {Machine review of arXiv:2508.21167}
}

read the original abstract

One in four people dementia live alone, leading family members to take on caregiving roles from a distance. Many researchers have developed remote monitoring solutions to lessen caregiving needs; however, limitations remain including privacy preserving solutions, activity recognition, and model generalizability to new users and environments. Structural vibration sensor systems are unobtrusive solutions that have been proven to accurately monitor human information, such as identification and activity recognition, in controlled settings by sensing surface vibrations generated by activities. However, when deploying in an end user's home, current solutions require a substantial amount of labeled data for accurate activity recognition. Our scalable solution adapts synthesized data from near-surface acoustic audio to pretrain a model and allows fine tuning with very limited data in order to create a robust framework for daily routine tracking.

Figures

Figures reproduced from arXiv: 2508.21167 by the authors.

**Figure 3.** RARR models. (a) The system first pre-trains a multitask VAE model on a large-scale ASMR audio dataset.(b) To adapt to the vibration modality, we introduce a low-rank 1D CNN adapter layer that finetunes the modality-specific latent parameters 𝜇 and 𝜎. In addition, we fine-tune the final linear layer of the temporal convolutional network (TCN) to account for modality-related distribution shifts. All other modules r… view at source ↗

**Figure 4.** Activity recognition accuracy of RARR and three baselines. various availability of data for different activities, we balance the dataset such that each activity label has an equal amount of samples. Home Health Simulation Suite Dataset. We collected realistic home activity induced vibration signal in the University of California Davis’ Home Health Simulation Suite. The Simulation Suite is a one-bedroom apartment t… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 18 canonical work pages

[1]

What Microphone is Used for ASMR? Exploring the Best Options for Optimal Tingles

2024. What Microphone is Used for ASMR? Exploring the Best Options for Optimal Tingles. https://blinksandbuttons.net/what-microphone- is-used-for-asmr [Online; accessed 14. Aug. 2025]

work page 2024
[2]

Ozan Baris, Yizhuo Chen, Gaofeng Dong, Liying Han, Tomoyoshi Kimura, Pengrui Quan, Ruijie Wang, Tianchen Wang, Tarek Abdelza- her, Mario Bergés, Paul Pu Liang, and Mani Srivastava. 2025. Founda- tion Models for CPS-IoT: Opportunities and Challenges. arXiv (Jan. 2025). https://doi.org/10.48550/arXiv.2501.16368 arXiv:2501.16368

arXiv doi:10.48550/arxiv.2501.16368 2025
[3]

Carl Doersch. 2016. Tutorial on Variational Autoencoders. arXiv (June 2016). https://doi.org/10.48550/arXiv.1606.05908 arXiv:1606.05908

arXiv doi:10.48550/arxiv.1606.05908 2016
[4]

Arindam Ghosh, Amartya Chakraborty, Dhruv Chakraborty, Mousumi Saha, and Sujoy Saha. 2023. UltraSense: A non-intrusive approach for human activity identification using heterogeneous ultrasonic sensor grid for smart home environment. Journal of Ambient Intelligence and Humanized Computing 14, 12 (2023), 15809–15830

work page 2023
[5]

Zhizhang Hu, Tong Yu, Yue Zhang, and Shijia Pan. 2020. Fine-grained activities recognition with coarse-grained labeled multi-modal data. In Adjunct Proceedings of the 2020 ACM International Joint Conference on Pervasive and Ubiquitous Computing and Proceedings of the 2020 ACM International Symposium on Wearable Computers . 644–649

work page 2020
[6]

Zhizhang Hu, Yue Zhang, Tong Yu, and Shijia Pan. 2022. VMA: Do- main Variance-and Modality-Aware Model Transfer for Fine-Grained Occupant Activity Recognition. In 2022 21st ACM/IEEE International Conference on Information Processing in Sensor Networks (IPSN) . IEEE, 259–270

work page 2022
[7]

Tomoyoshi Kimura, Jinyang Li, Tianshi Wang, Yizhuo Chen, Ruijie Wang, Denizhan Kara, Maggie Wigness, Joydeep Bhattacharyya, Mud- hakar Srivatsa, Shengzhong Liu, et al. 2024. Vibrofm: Towards micro foundation models for robust multimodal iot sensing. In2024 IEEE 21st International Conference on Mobile Ad-Hoc and Smart Systems (MASS) . IEEE, 10–18

work page 2024
[8]

Tomoyoshi Kimura, Jinyang Li, Tianshi Wang, Denizhan Kara, Yizhuo Chen, Yigong Hu, Ruijie Wang, Maggie Wigness, Shengzhong Liu, Mani Srivastava, et al . 2024. On the efficiency and robustness of vibration-based foundation models for iot sensing: A case study. In 2024 IEEE International Workshop on Foundation Models for Cyber- Physical Systems & Internet o...

work page 2024

Show all 20 references

[9]

Tomoyoshi Kimura, Ashitabh Misra, Yizhuo Chen, Denizhan Kara, Jinyang Li, Tianshi Wang, Ruijie Wang, Joydeep Bhattacharyya, Jae Kim, Prashant Shenoy, et al. 2024. The case for micro foundation mod- els to support robust edge intelligence. In 2024 IEEE 6th International Confere...

2024
[10]

Gierad Laput, Karan Ahuja, Mayank Goel, and Chris Harrison. 2018. Ubicoustics: Plug-and-play acoustic activity recognition. InProceedings of the 31st annual ACM symposium on user interface software and technology. 213–224

2018
[11]

Dawei Liang and Edison Thomaz. 2019. Audio-based activities of daily living (adl) recognition with large-scale acoustic embeddings from online videos. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 3, 1 (2019), 1–18

2019
[12]

Weixi Liang, Rongshan Tang, Sihan Jiang, Ruqi Wang, Yubin Zhao, and Cheng-Zhong Xu. 2023. LiWi-HAR: Lightweight WiFi-Based Human Activity Recognition Using Distributed AIoT. IEEE IoT J. 11, 1 (June 2023), 597–611. https://doi.org/10.1109/JIOT.2023.3286455

2023
[13]

Thomas S Murray, Daniel R Mendat, Kayode A Sanni, Philippe O Pouliquen, and Andreas G Andreou. 2017. Bio-inspired human action recognition with a micro-Doppler sonar system. IEEE Access 6 (2017), 28388–28403

2017
[14]

Viet Nguyen, Mohamed Ibrahim, Siddharth Rupavatharam, Minitha Jawahar, Marco Gruteser, and Richard Howard. 2018. Eyelight: Light- and-shadow-based occupancy estimation and room activity recogni- tion. In IEEE INFOCOM 2018-IEEE conference on computer communica- tions. IEEE, 351–359

2018
[15]

Shijia Pan, Ceferino Gabriel Ramirez, Mostafa Mirshekari, Jonathon Fagert, Albert Jin Chung, Chih Chi Hu, John Paul Shen, Hae Young Noh, and Pei Zhang. 2017. Surfacevibe: vibration-based tap & swipe tracking on ubiquitous surfaces. In 2017 16th ACM/IEEE International Conferenc...

2017
[16]

Risako Tanigawa and Yasunori Ishii. 2024. Hear-your-action: Human action recognition by ultrasound active sensing. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 7260–7264

2024
[17]

Hui Wei, Dong Yoon Lee, Shubham Rohal, Zhizhang Hu, Ryan Rossi, Shiwei Fang, and Shijia Pan. 2025. A Survey of Foundation Models for IoT: Taxonomy and Criteria-Based Analysis. arXiv (June 2025). https://doi.org/10.48550/arXiv.2506.12263 arXiv:2506.12263

2025 doi
[18]

Cunyi Yin, Jing Chen, Xiren Miao, Hao Jiang, and Deying Chen. 2021. Device-Free Human Activity Recognition with Low-Resolution In- frared Array Sensor Using Long Short-Term Memory Neural Net- work. Sensors (Basel). 21, 10 (May 2021), 3551. https://doi.org/10.3390/ s21103551 ar...

2021
[19]

Jia Zhang, Rui Xi, Yuan He, Yimiao Sun, Xiuzhen Guo, Weiguo Wang, Xin Na, Yunhao Liu, Zhenguo Shi, and Tao Gu. 2023. A survey of mmWave-based human sensing: Technology, platforms and applica- tions. IEEE Communications Surveys & Tutorials 25, 4 (2023), 2052– 2087

2023
[20]

Wenting Zhao, Shufeng Kong, Junwen Bai, Daniel Fink, and Carla Gomes. 2021. HOT-VAE: Learning High-Order Label Correlation for Multi-Label Classification via Attention-Based Variational Autoen- coders. AAAI 35, 17 (May 2021), 15016–15024. https://doi.org/10. 1609/aaai.v35i17.17762

2021

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

What Microphone is Used for ASMR? Exploring the Best Options for Optimal Tingles

2024. What Microphone is Used for ASMR? Exploring the Best Options for Optimal Tingles. https://blinksandbuttons.net/what-microphone- is-used-for-asmr [Online; accessed 14. Aug. 2025]

work page 2024

[2] [2]

Ozan Baris, Yizhuo Chen, Gaofeng Dong, Liying Han, Tomoyoshi Kimura, Pengrui Quan, Ruijie Wang, Tianchen Wang, Tarek Abdelza- her, Mario Bergés, Paul Pu Liang, and Mani Srivastava. 2025. Founda- tion Models for CPS-IoT: Opportunities and Challenges. arXiv (Jan. 2025). https://doi.org/10.48550/arXiv.2501.16368 arXiv:2501.16368

arXiv doi:10.48550/arxiv.2501.16368 2025

[3] [3]

Carl Doersch. 2016. Tutorial on Variational Autoencoders. arXiv (June 2016). https://doi.org/10.48550/arXiv.1606.05908 arXiv:1606.05908

arXiv doi:10.48550/arxiv.1606.05908 2016

[4] [4]

Arindam Ghosh, Amartya Chakraborty, Dhruv Chakraborty, Mousumi Saha, and Sujoy Saha. 2023. UltraSense: A non-intrusive approach for human activity identification using heterogeneous ultrasonic sensor grid for smart home environment. Journal of Ambient Intelligence and Humanized Computing 14, 12 (2023), 15809–15830

work page 2023

[5] [5]

Zhizhang Hu, Tong Yu, Yue Zhang, and Shijia Pan. 2020. Fine-grained activities recognition with coarse-grained labeled multi-modal data. In Adjunct Proceedings of the 2020 ACM International Joint Conference on Pervasive and Ubiquitous Computing and Proceedings of the 2020 ACM International Symposium on Wearable Computers . 644–649

work page 2020

[6] [6]

Zhizhang Hu, Yue Zhang, Tong Yu, and Shijia Pan. 2022. VMA: Do- main Variance-and Modality-Aware Model Transfer for Fine-Grained Occupant Activity Recognition. In 2022 21st ACM/IEEE International Conference on Information Processing in Sensor Networks (IPSN) . IEEE, 259–270

work page 2022

[7] [7]

Tomoyoshi Kimura, Jinyang Li, Tianshi Wang, Yizhuo Chen, Ruijie Wang, Denizhan Kara, Maggie Wigness, Joydeep Bhattacharyya, Mud- hakar Srivatsa, Shengzhong Liu, et al. 2024. Vibrofm: Towards micro foundation models for robust multimodal iot sensing. In2024 IEEE 21st International Conference on Mobile Ad-Hoc and Smart Systems (MASS) . IEEE, 10–18

work page 2024

[8] [8]

Tomoyoshi Kimura, Jinyang Li, Tianshi Wang, Denizhan Kara, Yizhuo Chen, Yigong Hu, Ruijie Wang, Maggie Wigness, Shengzhong Liu, Mani Srivastava, et al . 2024. On the efficiency and robustness of vibration-based foundation models for iot sensing: A case study. In 2024 IEEE International Workshop on Foundation Models for Cyber- Physical Systems & Internet o...

work page 2024