This paper introduces a Compatibility Prediction Latent World Model (LWM) for robot navigation. Existing world models usually reconstruct future observations or features to learn state transitions, which adds unnecessary complexity and offers limited decision‑making benefit. LWM instead predicts action‑conditioned latent feature compatibility, avoiding observation reconstruction. The key insight is that spatial proximity correlates with latent feature similarity, allowing direct evaluation of action consequences in latent space.
To enable counterfactual training, the model samples action sequences across trajectories and learns to predict which sequences bring the robot closer to the goal. This prediction capability lets the model supervise policy learning from unlabeled video data and further improve policies via reinforcement learning entirely within the latent world. The imagination‑driven framework thus eliminates the need for action annotations and extra environment interaction.
Extensive experiments on several real‑world robot navigation datasets show that LWM markedly outperforms prior world‑model and imitation‑learning methods in prediction accuracy, policy learning, and real‑world navigation performance. Code, pretrained models, and additional materials are publicly available.
Blogger's Review: By shifting focus from high‑dimensional observation reconstruction to latent feature compatibility, LWM achieves more efficient decision reasoning. The counterfactual sequence prediction and fully‑in‑model reinforcement learning are especially compelling, demonstrating a viable path to boost robot navigation under minimal labeling effort. This work offers valuable insights for future self‑supervised robotic learning.