Latent World Model Boosts Robot Navigation Policy Learning

This research introduces a novel approach to robot navigation using a compatibility prediction Latent World Model (LWM). Unlike traditional world models that reconstruct future observations, LWM focuses on predicting action-conditioned latent feature compatibility. The core idea is that spatial proximity correlates with latent feature similarity, allowing action consequences to be evaluated directly within a latent space, simplifying the prediction task. The model is designed to support counterfactual training by leveraging action sequences sampled across trajectories, learning to predict which sequences lead closer to a goal. This imagination-driven framework enables policy learning from unlabeled video data, eliminating the need for explicit action annotations or extensive environment interaction. Furthermore, the learned world model can supervise policy learning and refine policies through reinforcement learning entirely within its simulated environment. Extensive experiments conducted on multiple real-world robot navigation datasets demonstrate that this approach significantly surpasses existing world model and imitation learning methods. It shows improvements in prediction accuracy, policy learning efficiency, and actual real-world navigation performance, offering a promising direction for autonomous systems.