Predictive Visual Representations for Generalizable Vision-Based Navigation

This project investigates whether action-conditioned future latent prediction can improve visual representations learned by a vision-based navigation policy and their transfer to downstream reinforcement learning. The study uses the MetaDrive simulator to train and evaluate vision-based driving agents on procedurally generated road layouts, comparing behaviour cloning with and without auxiliary representation-learning objectives. The project includes controlled ablations of temporal prediction, action conditioning, and same-timestep cross-view invariance, together with downstream PPO transfer and representation analysis.