Researchers have released a large human activity recognition dataset that pools motion data from many devices, body positions and research projects into a single resource, according to a paper published in Scientific Data, a Nature portfolio journal. The AICOS-HAR dataset contains 5,231 labeled activity acquisitions totaling 101.4 hours, gathered from 106 participants using 87 smartphones and 17 wearable devices. Alongside the data, the authors published 21 open-source parsers that convert other public activity recognition datasets into one shared format.
The work is credited to Marília Barandas, Ana Cravidão Pereira, Carolina Carvalho, Duarte Folgado, Ricardo Santos, Maria Russo, Vânia Guimarães, Hugo Gamboa and Inês Sousa, with affiliations at the Comprehensive Health Research Center in Porto, Portugal, and the LIBPhys laboratory at NOVA University Lisbon.
Why it matters
Human activity recognition is the task of identifying what a person is doing — walking, sitting, climbing stairs — from motion signals picked up by everyday devices such as smartphones. Machine learning models handle much of this classification, and the field has advanced quickly. But according to the paper, progress is held back by fragmentation: publicly available datasets were built with different protocols, different hardware, different sensor placements and different labeling conventions.
That mismatch has a practical cost. A model that scores well on one dataset often performs poorly when it meets new people, new devices or messier real-world conditions, the authors write. The AICOS-HAR release is aimed squarely at that gap, offering both a broader pool of data and a way to compare models across datasets rather than within a single one.
The stakes extend beyond academic benchmarks. Activity recognition underpins health monitoring and mobility analysis, two application areas the authors single out. A model that only works on the hardware and population it was trained on has limited use in either setting.
What is inside the dataset
The collection spans multiple research initiatives rather than a single controlled study, which is part of what gives it range. It covers inertial signals and barometric signals — the latter useful for detecting changes in altitude, such as moving between floors. Data was captured across a variety of body positions and device types, reflecting the fact that people carry phones in pockets, hands or bags, and wear sensors in different places.
The authors note that earlier versions of this dataset have already been used in published work on position-invariant activity recognition, pedestrian navigation and active learning. That track record is presented as evidence of the resource's usefulness rather than a claim about the new release's performance.
The 21 parsers may prove just as consequential as the raw data. Each one takes an existing public activity recognition dataset and rewrites it into a unified schema, which allows researchers to combine sources and run systematic cross-dataset benchmarks. In effect, the parsers turn a scattered collection of incompatible files into something that can be analyzed as a whole.
Funding and access
The paper states that the work received partial support from European funds through the Recovery and Resilience Plan, under the project "Center for Responsible AI," project number C645008882-00000055. Publication was supported by Fundação para a Ciência e Tecnologia (FCT-Portugal) through national funds, within the scope of UID/06291/2025.
The article is open access under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits non-commercial sharing and redistribution with attribution but does not allow distribution of adapted material. The dataset and parsers are described as open source. The full citation is Barandas, M., Pereira, A.C., Carvalho, C. et al., Scientific Data (2026), DOI 10.1038/s41597-026-08470-y.
Springer Nature, the publisher, states that it remains neutral regarding jurisdictional claims in published maps and institutional affiliations.
What to watch next is whether the parsers succeed in doing for the field what the dataset alone cannot: making results from different research groups directly comparable, and showing which models actually hold up when the device, the wearer and the day all change.
