What is it about?

Smartwatches, fitness trackers, smartphones, and other wearable devices continuously collect sensor data that can be used to understand human activities such as walking, running, exercising, or sleeping. This capability, known as Human Activity Recognition (HAR), is becoming increasingly important for healthcare, fitness, rehabilitation, workplace safety, and many other applications. However, current activity recognition systems often struggle when they are used by new people, on different devices, or in unfamiliar environments. They also typically require large amounts of labeled training data, which can be expensive and time-consuming to collect. Recent advances in artificial intelligence have introduced a new generation of large-scale AI models, often called “foundation models.” Instead of being trained for a single task, these models learn general knowledge from massive amounts of data and can then be adapted to many different applications. Similar approaches are beginning to transform sensor-based activity recognition. In this survey, we review and organize the rapidly growing research area of foundation models for wearable and sensor-based activity recognition. We explain how these models are built, how they learn from large collections of sensor data, and how they can be adapted to recognize activities in real-world situations. We also identify major research trends, discuss current challenges such as privacy and personalization, and highlight opportunities for creating more reliable, adaptable, and human-centered activity recognition systems. Our goal is to provide researchers, students, and practitioners with a clear overview of this emerging field and to help guide future developments in wearable and ubiquitous computing. The repo for this emerging topic is: https://github.com/zhaxidele/Foundation-Models-Defining-A-New-Era-In-Human-Activity-Recognition

Featured Image

Why is it important?

Human Activity Recognition is a key technology behind many wearable and ubiquitous computing applications, including health monitoring, fitness tracking, rehabilitation, elderly care, and smart environments. As sensor-equipped devices become increasingly common, there is growing demand for AI systems that can reliably understand human behavior across different users, devices, and real-world conditions. Recent advances in large-scale AI models have the potential to significantly improve the robustness, adaptability, and scalability of activity recognition systems. However, research in this area is evolving rapidly, making it difficult for researchers and practitioners to understand the available approaches, their strengths, and their limitations. This survey provides the first comprehensive overview of foundation models for sensor-based activity recognition. By organizing existing knowledge, identifying emerging trends, and highlighting open challenges, it helps accelerate future research and supports the development of more accurate, personalized, privacy-preserving, and accessible activity-aware technologies that can benefit everyday life.

Perspectives

From my perspective, this publication arrives at an important turning point for the Human Activity Recognition community. For many years, progress in HAR has largely been driven by designing task-specific models and evaluating them on individual datasets. Recently, however, we have started to see a shift toward large-scale pretrained models that can transfer knowledge across tasks, users, devices, and sensing modalities. While foundation models have already transformed fields such as natural language processing and computer vision, their role in sensor-based activity recognition remains less clearly understood. Through this survey, I wanted not only to summarize existing work, but also to provide a structured framework for thinking about where the field is heading. In particular, I believe future HAR systems will move beyond isolated recognition models toward more general, adaptable, and human-centered AI systems that can continuously learn, personalize, and collaborate with users. I also hope this survey serves as a useful entry point for researchers from neighboring communities, helping to connect advances in wearable sensing, time-series learning, multimodal AI, and large language models. Ultimately, the goal is to encourage a broader discussion about how foundation models can enable the next generation of intelligent, trustworthy, and widely deployable activity-aware systems.

Sizhen Bian
Northwestern Polytechnical University

Read the Original

This page is a summary of: Foundation Models Defining A New Era In Sensor-based Human Activity Recognition: A Survey And Outlook, Proceedings of the ACM on Interactive Mobile Wearable and Ubiquitous Technologies, June 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3810230.
You can read the full text:

Read

Resources

Contributors

The following have contributed to this page