"Factorised Latent Dynamics for Visual Navigation in World Models". Corelatore: Amit Ranjan Trivedi. Dreamer-style world models for visual control usually maintain the history of interaction through recurrent models of the latent state based on gated dynamics. This thesis studies whether visual navigation with sparse rewards can benefit from a more structured latent transition, inspired by the Tolman-Eichenbaum Machine (TEM), a model of spatial memory that separates path integration, sensory information, grounded state representations, and associative memory. The proposed TEM-style agent is inserted into a simplified, controlled Dreamer-style scaffold and compared with a corresponding RSSM-GRU baseline. Both agents share obser- vations, replay buffer structure, predictive heads, actor-critic training procedure, checkpoint selection protocol, and validation and held-out test seeds.