Physical Self-Supervised Learning Improves IMU Sensing.

The research presents a novel paradigm called physical self-supervised learning, designed to overcome the limitations of costly labeled data and poor robustness in IMU-based sensing. Unlike existing methods that still require some labeled data for domain adaptation, this approach eliminates the need for manual labels entirely. The core innovation lies in replacing the conventional neural decoder with an auto-adaptive physics decoder. This decoder is a learnable family of kinematic equations that not only enforces explicit physical structure but also adapts across diverse environments. The framework also incorporates a hybrid two-stage IMU encoder with reconstruction in a structured latent space to mitigate sensor noise, probabilistic frequency-spatial constraints to disentangle motion, a multi-view kinematic tree for sparse signals, and an uncertainty-aware formulation to handle inference ambiguity. This results in substantial error reductions for inertial tracking and motion capture across challenging generalization scenarios.