EdgeHAR: An Edge-Native Compact Sensor Foundation Model for Human Activity Recognition

He Zhang; Siyu Yuan; Siyu Liu; Sizhen Bian; Bin Guo

cs.LGSep 13, 2026

EdgeHAR: An Edge-Native Compact Sensor Foundation Model for Human Activity Recognition

Authors: He Zhang, Siyu Yuan, Siyu Liu, Sizhen Bian, Bin Guo

Organizations: Northwestern Polytechnical University Xi’an, Shaanxi, China · RPTU University Kaiserslautern-Landau Kaiserslautern, Rheinland Pfalz, Germany · Xi’an University of Technology Xi’an, Shaanxi, China

Abstract

Sensor-based human activity recognition (HAR) is fundamental to ubiquitous and wearable computing, yet existing foundation models are largely designed for cloud-scale deployment and struggle with real-world sensing shifts, including unseen users, devices, sampling rates, and sensor placements. We present EdgeHAR, an edge-native compact sensor foundation model designed for wearable intelligence. Unlike conventional models that entangle activity knowledge with acquisition variations, EdgeHAR learns transferable representations by factorizing sensor signals into three latent codes: an (i)Activity-Semantic Code capturing reusable activity knowledge, a (ii)Motion-Dynamics Code modeling temporal patterns, and an (iii)Acquisition-Context Code representing sensor-specific variations. This disentangled design enables efficient adaptation to new users, devices, placements, and activity classes with limited target-domain data. By incorporating lightweight adaptation modules, EdgeHAR achieves foundation-model-level transferability while satisfying edge constraints in computation, memory, latency, and privacy. Experiments across heterogeneous HAR datasets demonstrate that EdgeHAR maintains competitive recognition performance under distribution shifts with substantially reduced deployment cost. EdgeHAR establishes a practical paradigm for compact, edge-first foundation models for ubiquitous sensing systems.

Figures & tables

Dataset	Subjects	Sensors	Position	Rate	Classes
HHAR ( Stisen et al., 2015 )	9	acc+gyro	Multiple	50 HZ	6
MotionSense ( Malekzadeh et al., 2019 )	24	acc+gyro	Pocket	50 HZ	6
UCI-HAR ( Reyes-Ortiz et al., 2016 )	30	acc+gyro	Waist	50 HZ	6
Shoaib ( Shoaib et al., 2014 )	10	acc+gyro+mag	Multiple	50 HZ	7
PAMAP2 ( Reiss and Stricker, 2012 )	9	acc+gyro+mag	Multiple	100 HZ	18

Table 1. Statistics of wearable sensing datasets used for EdgeHAR pretraining and evaluation.

Source	Target	SDMix	Contrast -Sense	CrossHAR	OUR-ZS	OUR-FS
S+M+H	UCI	75.38	79.43	88.71	62.85	92.15
U+S+H	Motion	64.16	53.20	76.12	54.23	88.68
H+U+M	Shoaib	70.11	63.83	73.52	58.62	86.28
S+U+M	HHAR	71.10	71.75	74.36	47.12	92.45

Table 2. Multi-source cross-dataset transfer performance (ACC%). ZS is zero-short and FS is few-short.

Source	Target	Transformer	UniHAR	LLM4HAR	OUR-ZS	OUR-FS
UCI	Shoaib	25.99	69.90	74.98	63.32	90.70
	Motion	35.64	76.08	87.09	62.45	88.23
	HHAR	44.52	75.43	85.75	61.08	96.08
Motion	UCI	39.36	69.05	89.44	67.96	93.09
	Shoaib	44.26	78.71	73.70	77.64	93.01
	HHAR	42.97	58.33	75.21	79.68	97.53

Table 3. Single-source cross-dataset transfer performance (ACC%). ZS is zero-short and FS is few-short.

Method	UCI	Motion	HHAR	Avg.
SDMix	90.33	75.48	77.44	81.08
ContrastSense	93.33	83.13	87.66	88.04
DI2SDiff	87.68	88.43	89.19	88.43
OUR-ZS	86.19	85.42	88.97	86.86
OUR-FS	91.88	93.17	94.89	93.31

Table 4. Cross-person transfer performance (ACC%) on unseen users.

Source	Target	Mixup	ContrastSense	DI2SDiff	OUR-ZS	OUR-FS
B+P+W	Arm	40.17	70.81	74.14	6.00	95.71
A+P+W	Belt	39.87	68.06	68.71	19.90	88.29
A+B+W	Pocket	47.01	71.48	72.36	69.25	90.19
A+B+P	Wrist	41.12	75.54	77.11	42.09	91.59

Table 5. Cross-position transfer performance (ACC%) on shoaib dataset. ZS is zero-short and FS is few-short.

Model	Params(M)	Size(MB)	CPU(ms)	Memory(MB)
GPT-2 Small	160.0	400.0	1028.0	–
LLMHAR	8.0	46	104.5	–
CNN	6.45	25.805	2.14	51.05
DeepConvLSTM	0.46	1.85	40.90	10.16
HART	1.45	5.91	5.38	12.74
MobileHART	2.54	10.35	13.49	20.77

Table 6. Efficiency comparison under edge deployment.

Emerging Activity	K=20	K=50	K=100
Lying back	0.0022	0.6674	0.8611
Walking parking lot	0.3260	0.8871	0.9899
Cycling bike horizontal	0.0596	0.3849	0.9413
Playing basketball	0.2471	0.6147	0.8786
Macro-F1	0.1627	0.7487	0.9186
Overall Accuracy	0.2382	0.7574	0.9190

Table 7. Few-shot adaptation to emerging activities on PAMAP2. K denotes the number of target samples per activity.

Panel of Figure 2 — Figure 2. Effect of target-domain fine-tuning data ratio on adaptation performance. EdgeHAR is evaluated under different amounts of target data across heterogeneous datasets.

Explore similar work

RAG-HAR+: Towards Cost-Efficient LLM-Based Human Activity Recognition for Edge Deployment
Jul 29, 2026Hansi Karunarathna, Nirhoshan Sivaroopan, Chamara Madarasingha +2Human Activity Recognition Cost-Aware Inference
University of Sydney, Australia · Curtin University, Australia · Colorado State University, USA
STELLA: Efficient Sensor-to-LLM Translation for On-Device Human Activity Recognition
Jul 3, 2026Nirhoshan Sivaroopan, Albert Zomaya, Kanchana ThilakarathnaOn-Device Language Model Inference Human Activity Recognition
The University of Sydney · Australia
GenHAR: Generalizing Cross-domain Human Activity Recognition for Last-mile Delivery
May 21, 2026Zhiqing Hong, Zelong Li, Xiubin Fan +5Human Activity Recognition Domain Generalization
HKUST (GZ) Guangzhou, China · JD Logistics Beijing, China · Beijing Normal–Hong Kong Baptist University Zhuhai, China +2

Jul 29, 2026J/K move · Enter open · S save

RAG-HAR+: Towards Cost-Efficient LLM-Based Human Activity Recognition for Edge Deployment

Hansi Karunarathna, Nirhoshan Sivaroopan, Chamara Madarasingha, Anura Jayasumana+1

University of Sydney, Australia · Curtin University, Australia · Colorado State University, USA

Human Activity Recognition Cost-Aware Inference Edge ML Retrieval-Augmented ICL

Open paper AskPDF arXiv

Human Activity Recognition (HAR) from wearable sensors supports applications in healthcare, rehabilitation, fitness tracking, and smart environments. Yet, existing deep learning approaches require dataset-specific training, large labeled corpora, and repeated adaptation to new sensor settings or activity taxonomies. Retrieval-Augmented Generation for Human Activity Recognition (RAG-HAR) addresses this by framing HAR as a training-free, retrieval-augmented task, in which statistical descriptions of sensor windows are used to retrieve similar labeled examples that guide LLM-based classification. We introduce RAG-HAR+, a retrieval-first and cost-optimized extension that strengthens retrieval while reducing dependence on LLM-based inference. RAG-HAR+ uses an offline Retrieval Designer Agent to design dataset-specific feature groups from a diverse pool of motion descriptors, enabling sensor windows to be compared using features better aligned with dataset-specific activity patterns. During inference, RAG-HAR+ uses majority voting over retrieved neighbors for samples with strong retrieval evidence and defers only uncertain cases to an LLM-based Ambiguity Resolver Agent. Across six HAR benchmarks, RAG-HAR+ maintains competitive or improved performance while reducing LLM usage, token consumption, and inference time. We further extend the RAG-HAR mobile prototype to demonstrate the practical feasibility of retrieval-first, LLM-assisted HAR in mobile sensing scenarios.

Abstract

Explore similar work

RAG-HAR+: Towards Cost-Efficient LLM-Based Human Activity Recognition for Edge Deployment

STELLA: Efficient Sensor-to-LLM Translation for On-Device Human Activity Recognition

GenHAR: Generalizing Cross-domain Human Activity Recognition for Last-mile Delivery