Signal MyoMechanix: Biomechanically-Grounded Compositional Skilled Activity Understanding and Coaching

Summary

Researchers introduce MyoMechanix, a multimodal ecosystem for weight-loaded physical actions that aligns motion data with muscle activity, addressing the limitation that existing action-quality-assessment datasets rely mainly on visual inputs and overlook physiological dynamics. The expert-annotated dataset contains more than 7,500 samples across 20 actions from 38 subjects, with synchronized multiview RGB video, 3D pose, surface EMG, and other physiological signals, which the authors describe as the largest multimodal action-quality-assessment benchmark to date. They also build a Fitness Knowledge Graph organizing expert annotations into structured relationships among actions, phases, errors, and corrective feedback, and develop CUBIST, a reasoning engine for fine-grained error attribution that achieves state-of-the-art results. The work targets 'Physical AI' applications in fitness, rehabilitation, and healthcare. The paper was submitted to arXiv's cs.AI category on August 26, 2026.