The perception half of active inference is the variational autoencoder loss under another name, so it adds no new learning objective.

The claim

Mazzaglia, Verbelen, Çatal and Dhoedt (2022) point out that the variational free energy an active inference agent minimises over past experience is, in machine-learning terms, the negative evidence lower bound, the same loss used to train variational autoencoders. So the perception half of active inference adds no new objective to machine learning; whatever the principle contributes for artificial agents has to come from elsewhere (action selection, priors, or the unifying story).