MWM: Mobile World Models for Action-Conditioned Consistent Prediction
This work improves the reliability and efficiency of embodied navigation systems for robots and autonomous agents by enhancing the consistency of future state predictions and enabling faster inference.
This paper introduces MWM, a mobile world model designed for planning-based image-goal navigation. It addresses the issue of action-conditioned inconsistency in existing navigation world models, which leads to prediction drift and degraded planning, and also tackles the challenge of efficient deployment with few-step diffusion inference while preserving rollout consistency. The proposed MWM demonstrates consistent gains across visual fidelity, trajectory accuracy, planning success, and inference efficiency on both benchmark and real-world tasks.
World models enable planning in imagined future predicted space, offering a promising framework for embodied navigation. However, existing navigation world models often lack action-conditioned consistency, so visually plausible predictions can still drift under multi-step rollout and degrade planning. Moreover, efficient deployment requires few-step diffusion inference, but existing distillation methods do not explicitly preserve rollout consistency, creating a training-inference mismatch. To address these challenges, we propose MWM, a mobile world model for planning-based image-goal navigation. Specifically, we introduce a two-stage training framework that combines structure pretraining with Action-Conditioned Consistency (ACC) post-training to improve action-conditioned rollout consistency. We further introduce Inference-Consistent State Distillation (ICSD) for few-step diffusion distillation with improved rollout consistency. Our experiments on benchmark and real-world tasks demonstrate consistent gains in visual fidelity, trajectory accuracy, planning success, and inference efficiency. Code: https://github.com/AIGeeksGroup/MWM. Website: https://aigeeksgroup.github.io/MWM.