Bridging Skeletal Motion and Wearable Sensing: A Survey of LLM-driven Human Motion Computing

ACL ARR 2026 January Submission8622 Authors

Keywords: LLM, Human Motion, Skeletal, Wearable Sensor, Multimodal Learning, Human Activity Recognition

Abstract: Human motion computing is experiencing a paradigm shift from traditional deep learning approaches to large language model (LLM)-driven frameworks. Following the PRISMA guideline, this paper systematically analyzed 101 core papers and 21 multimodal datasets. To bridge the gap between skeletal motion sequences and wearable sensor data, we propose a full-stack classification framework encompassing motion representation, modeling, and agent interaction. We examine how LLMs enhance motion perception, understanding, generation, and planning, and summarize emerging application paradigms in advanced domains such as healthcare, sports science, and embodied intelligence. Finally, we proposed future directions for bottlenecks such as reasoning delay, interpretability, and multi-modal unification, aiming to drive the leap of human motion computing from single sequence fitting to general motion intelligent systems with physical feedback and long-term planning capabilities.

Paper Type: Long

Research Area: Multimodality and Language Grounding to Vision, Robotics and Beyond

Research Area Keywords: multimodality,cross-modal application,ross-modal content generation,cross-modal pretraining,cross-modal machine translation

Contribution Types: Surveys

Languages Studied: English

Submission Number: 8622