CBA lab at Ubicomp / ISWC 2026
The CBA lab will attend and participate in Ubicomp / ISWC 2026, held from October 11–15, 2026 in Shanghai, China. This year, we are presenting three ISWC/Workshop publications. Thomas is also giving a number of talks and will be on panels as listed below. Come find us at the conference and let’s have a chat on all things Ubicomp and how we can work together.
Overview
- 3 papers
- 3 Keynotes
- 1 Expert Talk at Doctoral Colloquium (TP)
- 1 Panel Discussion
- PACM IMWUT editorial board (TP)
Papers
đź“„ On Self-Supervised Learning Using Virtual IMU data for Sensor-Based Human Activity Recognition
Authors
Elizabeth Bruda, Thomas Ploetz, Harish Haresamudram (Georgia Institute of Technology)
Workshop on “Ubiquitous Connectivity for Agentic and Physical AI”
đź“„ Read PDF
Summary:
Virtual sensor data generation has emerged as an effective avenue for tackling the `small labeled dataset problem’ in sensor-based human activity recognition, with the ability to generate large quantities of data without additional human data collection.
Self-supervised learning is an alternative approach, relying on existing large-scale unlabeled datasets for learning useful representations, which scales well with increasing data size and diversity.
In this paper, we investigate whether a combination of these two approaches, i.e., performing self-supervised pre-training on large-scale virtual sensor data (over 300 hours duration from 80 activities), is effective at tackling the small labeled dataset problem.
Surprisingly, we discover that virtual data alone cannot function as a replacement for data collected from human participants, and self-supervised methods are unable to effectively leverage the increased scale and diversity of sensor data.
We investigate the impact of two measures that tackle this shortcoming: data calibration, and utilizing a mixture of real and virtual data for model development.
Both measures improve performance, while falling short of supervised training incorporating the measures.
These findings pave the way for deeper examination of virtual data generation by the wearables community, and aid the development of methods that can tackle the challenges associated with such data.
đź“„ Models Got Talent: Identifying High Performing Wearable Human Activity Recognition Models Without Training
Authors
Richard Goldman, Varun Komperla, Thomas Ploetz, Harish Haresamudram (Georgia Institute of Technology)
ISWC
đź“„ Read PDF
Summary:
Wearable computing was originally imagined as computation woven into what people already wear, yet HAR has largely advanced through rigid or semi-rigid devices such as smartphones, smartwatches, and strapped IMUs, where sensor motion closely tracks body motion. Smart garments revisit this clothing-integrated vision while challenging this assumption, since fabric-mounted sensors also capture garment motion, fit variation, and sensor displacement. We present \textit{StitchSense}, a multi-IMU smart vest for studying garment-mounted activity recognition and transfer to wrist-worn HAR. StitchSense places five IMUs at the shoulders, chest, and hips to capture upper-body, trunk, and locomotion-related movement. We evaluate it both as a direct garment-based HAR platform and as a source domain for watch-based recognition under limited target calibration. Results show that StitchSense captures coarse functional movement structure, but standalone fine-grained recognition remains challenging due to garment motion, fit variation, sensor instability, and participant variability. In transfer experiments, combining StitchSense source data with limited target-watch calibration improves watch-based recognition, suggesting that smart garments can complement rather than replace wrist-worn devices.
đź“„ StitchSense: Smart-Garment Sensing as a Source Domain for Watch-Based Activity Recognition
Authors
Ishita Datta, Zikang Leng, Rachit Bhayana, Thomas Ploetz (Georgia Institute of Technology)
ISWC
đź“„ Read PDF
Summary:
Discovering high performing model architectures for wearables-based Human Activity Recognition (HAR) applications is challenging. The astonishing diversity and variability due to differing sensor locations, recording apparatus, activities, etc., can cause established architectures to perform worse on datasets/tasks they were not designed for. A promising complement to Neural Architecture Search (NAS) involves the development of Zero Cost Proxies (ZCPs), which correlate well with trained performance, but can be computed through a single forward/backward pass on a randomly sampled batch of data. In this paper, we investigate the effectiveness of eight ZCPs on six benchmark HAR datasets, and demonstrate that the top-predicted architectures obtain performance within 7% of that attained by full-scale training of 2,000 randomly sampled architectures. Furthermore, training the top-10 predicted architectures results in performance within 2% of full-scale training, leading to substantial computational savings. Our experiments introduce ZCPs to sensor-based HAR and demonstrate their suitability as an addition to NAS pipelines in practical scenarios
Expert Talk at DC
Thomas will be interviewed in an expert talk at the Doctoral Colloquium
Title
“Expert talk on academic publishing, with a focus on IMWUT.”
đź”—
Date and Time
October 11, 2026 at 4:00pm
Location
Room 5J
Panel at the tutorial on “Prototyping Your Personal LLM Health Agent”
Thomas will be panelist at the closing panel of the tutorial
Title
TBD
đź”—
“Prototyping Your Personal LLM Health Agent”
Date and Time
October 12, 2026 at 4:45pm
Location
Room 5A
Keynotes
Thomas will speak at “Reproduce: The First Workshop on Reproducible Methods for Wearable Sensing and Ubiquitous Computing”
Title
“TBD”
Abstract
TBD
Link
Date and Time
October 11, 2026 at 8:45am
Location
Room 5H
Thomas will speak at “UbiSense: Emerging Techniques in Ubiquitous Sensing and Interaction”
Title
“Agents Up Our Sleeves: From Passive Sensing to Autonomous Activity Recognition using (Agentic) AI”
Abstract
For decades, research into sensor-based Human Activity Recognition (HAR) has been a cornerstone of mobile and ubiquitous computing—evolving from early proofs of concept to integration across diverse application domains. While deep learning has rendered many traditional challenges manageable—leading some to claim HAR is a “solved problem”—real-world deployments still struggle with performance bottlenecks and the rigid, passive nature of current systems. In the era of modern AI, two of the most persistent bottlenecks in sensor-based HAR can finally be addressed head-on. First, data scarcity is being overcome through cross-modality transfer, self-supervised representation learning, and synthetic data generation. Second, the rise of agentic AI is enabling a shift from reactive pipelines—which merely map continuous sensor streams to static, predefined labels—to dynamic, open-ended activity recognition. These agentic workflows allow systems to autonomously adapt to shifting requirements, emerging target activities, and varying sensor availability. This keynote serves two purposes. First, I will offer a critical overview of where sensor-based HAR stands today, focusing on the practical challenges of deploying robust systems in complex scenarios. Second, I will examine the promises and pitfalls of agentic AI workflows in this domain. Drawing from emerging research, I will outline both a wishlist and a research roadmap for the next generation of sensor-based HAR systems that seamlessly integrate into practical applications.
Link
“UbiSense: Emerging Techniques in Ubiquitous Sensing and Interaction”
Date and Time
October 12, 2026 at 9:00am
Location
Room 5H
Thomas will speak at the tutorial on “Prototyping Your Personal LLM Health Agent”
Title
TBD
Abstract
TBD
Link
“Prototyping Your Personal LLM Health Agent”
Date and Time
October 12, 2026 at 2:00pm
Location
Room 5A
We look forward to engaging with the community at Ubicomp / ISWC 2026—see you there!