Embodied Predictive Coding explores linking a model's state predictions to actions in simple control tasks. The experiments use predictive-coding updates to reduce differences between expected and observed state, with actions connected to those updates.
Paddle and pendulum tasks provide small environments for examining the behavior. Recurrent and self-predicting hidden-state variants look at how internal state can persist while the model continues to learn.
One unresolved issue is settling into stable but unhelpful states. Reducing an internal prediction error does not by itself establish that the task is being solved. The work remains an experiment in the relationship between prediction, state dynamics, and control.