How would you design a self-supervised pretraining objective for a model that needs to predict both future video frames and vehicle control actions?
