MWM: Mobile World Models for Action-Conditioned Consistent Prediction

2026-03-08

AI Summary

This work addresses the prediction drift in existing navigation world models during multi-step rollouts, caused by inconsistent action conditioning, as well as the mismatch between training and few-step diffusion inference that degrades planning performance and deployment efficiency. To tackle these issues, the authors propose MWM, a novel world model featuring a two-stage training framework: first, structural pretraining, followed by post-training with Action Conditioning Consistency (ACC)—a mechanism explicitly integrated into the world model for the first time to enhance multi-step prediction consistency. Additionally, they introduce Inference-Consistent State Distillation (ICSD) to bridge the gap between full-step training and few-step inference. Experiments demonstrate that MWM significantly outperforms baseline methods in visual fidelity, trajectory accuracy, planning success rate, and inference efficiency.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Multiagent Systems: Multiagent Planning
• Planning, Routing, and Scheduling: Planning with Language Models

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

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.

Problem

• world models
• action-conditioned consistency
• rollout drift
• diffusion distillation
• embodied navigation

Innovation

• World Models
• Action-Conditioned Consistency
• Diffusion Distillation
• Inference-Consistent State Distillation
• Embodied Navigation