Developing World Models from Wearable Sensors

Wearable foundation models have demonstrated strong representation learning across heterogeneous physiological sensors, but largely remain focused on encoding observations rather than modeling how health states evolve over time. We introduce SensorWM, an action-conditioned latent world model for longitudinal wearable sensing that learns physiological transition dynamics from large-scale state-behavior trajectories. SensorWM models second-order dynamics in a compact latent space, enabling autoregressive multi-week forecasting and hypothetical behavioral rollouts without reconstructing raw observations. Across wearable forecasting tasks, the model outperforms persistence-based and SOTA time-series baselines while preserving clinically informative trajectory structure. We further demonstrate how action-conditioned rollouts can simulate alternative behavioral trajectories and translate them into interpretable changes in health. These results suggest that modeling physiological dynamics provides a promising direction for wearable intelligence.