MWM: Mobile World Models for Action-Conditioned Consistent Prediction

Han Yan; Zishang Xiang; Zeyu Zhang; Hao Tang

cs.CVMar 8, 2026

MWM: Mobile World Models for Action-Conditioned Consistent Prediction

Authors: Han Yan, Zishang Xiang, Zeyu Zhang, Hao Tang

Organizations: School of Computer Science, Peking University

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.

Figures & tables

Panel of Figure 1 — Fig. 2 : Overview of the Two-stage training pipeline for MWM. Our training paradigm first performs structure pretraining to learn fine-grained geometry and illumination-dependent appearance, then applies ACC post-training to mitigate compounding error while freezing the CDiT backbone and updating only AdaLN. Within post-training, we introduce ICSD to enable distillation that preserves the consistency objective, while aligning truncated training-time estimates with the inference-time endpoint.

Panel of Figure 2 — Fig. 3 : Qualitative results on SCAND. The predicted frames exhibit action-conditioned consistency (ACC) with the ground-truth frames.

Model	1s		2s		4s		8s		16s
	LPIPS $\downarrow$	DreamSim $\downarrow$	LPIPS $\downarrow$	DreamSim $\downarrow$	LPIPS $\downarrow$	DreamSim $\downarrow$	LPIPS $\downarrow$	DreamSim $\downarrow$	LPIPS $\downarrow$	DreamSim $\downarrow$
NWM (DDIM 25) [ 30 ]	0.478	0.309	0.494	0.308	0.508	0.321	0.540	0.345	0.569	0.373
NWM (DDIM 5) [ 30 ]	0.581	0.436	0.591	0.435	0.614	0.455	0.678	0.513	0.734	0.568
MWM (ours, DDIM 5)	0.368	0.244	0.395	0.260	0.421	0.276	0.459	0.306	0.495	0.337

TABLE I : ACC measured by LPIPS and DreamSim at different rollout horizons on SCAND.

Model	FID $\downarrow$
	1s	2s	4s	8s	16s
NWM (DDIM 25) [ 30 ]	96.68	91.57	90.37	91.29	93.63
NWM (DDIM 5) [ 30 ]	167.43	153.87	148.70	158.13	168.28
MWM (ours, DDIM 5)	80.97	87.92	85.19	85.80	93.12

TABLE II : Distribution-level visual fidelity on SCAND.

Model	Average Rollout Time (s) $\downarrow$
NWM (DDIM 25) [ 30 ]	9.6
NWM (DDIM 5) [ 30 ]	2.6
MWM (ours, DDIM 5)	2.3

TABLE III : Inference efficiency measured on an NVIDIA RTX PRO 6000 GPU.

Model	ATE $\downarrow$	RPE $\downarrow$
Forward	2.97	0.62
GNM [ 22 ]	2.12	0.61
NoMaD [ 24 ]	2.24	0.49
NWM + NoMaD ( $\times$ 16) [ 30 ]	2.18	0.48
NWM + NoMaD ( $\times$ 32) [ 30 ]	2.19	0.47
NWM [ 30 ]	1.28	0.33

TABLE IV : Goal-image navigation performance on SCAND with CEM planning.

Fig. 4 : Qualitative real-world evaluation . MWM generates planned rollouts that align better with real observations than NWM, indicating reduced multi-step error accumulation and improved planning quality. Blank frames in the figure indicate cases where the robot was forcibly emergency-stopped due to imminent collision.

Loss	FID (16s) $\downarrow$	DreamSim (16s) $\downarrow$	LPIPS (16s) $\downarrow$
L2 loss	352.62	0.769	0.883
L1 loss	341.77	0.750	0.857
LPIPS loss (ACC)	93.12	0.337	0.495

TABLE V : Ablation of ACC loss on SCAND.

Training paradigm	FID (16s) $\downarrow$	DreamSim (16s) $\downarrow$	LPIPS (16s) $\downarrow$
Only structure training	119.00	0.480	0.715
Only ACC training	196.72	0.657	0.729
Structure training + ACC training	98.17	0.382	0.639

TABLE VI : Ablation of training paradigm on MMK2-RealNav.

Rollout context in ACC training	FID (16s) $\downarrow$	DreamSim (16s) $\downarrow$	LPIPS (16s) $\downarrow$
$\hat{s}_{<\tau}^{(0)}$	109.75	0.400	0.569
$s_{<\tau}^{IC}$ (ICSD)	93.12	0.337	0.495

TABLE VII : Ablation of ICSD on SCAND.

Fig. 5 : Real-world deployment setup on the AIRBOT Mobile Manipulation Kit 2 (MMK2). (a) Hardware platform. (b) Deployment process.

Model	SR $\uparrow$	NE $\downarrow$
NoMaD [ 24 ]	0.08	2.88
NWM [ 30 ]	0.20	3.18
MWM (ours)	0.30	2.16

TABLE VIII : Quantitative real-world goal-image navigation results aggregated over four targets.

World Models for Robotics

World models for robotics predict how scenes and robot states change under actions to support planning, control, and learning.

See all 369 papers

G $^2$ WAM: World Action Models for Efficient and Geometry Grounded Image-Goal Navigation
Oct 12, 2026Ke Gui, Chenghao Shi, Ziyue Wang +3Scene Coordinate Regression Image-Goal Navigation
College of Intelligence Science and Technology, National University of Defense Technology, China · National Key Laboratory of Equipment State Sensing and Smart Support, National University of Defense Technology, China · School of Future Technology, National University of Defense Technology, China
SkipWorld: Latent Skip Planning with Distilled World Models for Efficient Visual Navigation
Oct 13, 2026Yifei Yang, Zehua Fan, Zike Yan +2Visual Navigation World Model-Based Planning
Institute for AI Industry Research (AIR), Tsinghua University · Hong Kong Embodied AI Lab · The Chinese University of Hong Kong +1
NavWAM: A Navigation World Action Model for Goal-Conditioned Visual Navigation
Jun 11, 2026Daichi Azuma, Taiki Miyanishi, Koya Sakamoto +6World Models for Robotics Robot Navigation
The University of Tokyo · National Institute of Informatics · AIRoA +1

Oct 12, 2026J/K move · Enter open · S save

G $^2$ WAM: World Action Models for Efficient and Geometry Grounded Image-Goal Navigation

Ke Gui, Chenghao Shi, Ziyue Wang, Longwen Chen+2

College of Intelligence Science and Technology, National University of Defense Technology, China · National Key Laboratory of Equipment State Sensing and Smart Support, National University of Defense Technology, China · School of Future Technology, National University of Defense Technology, China

Scene Coordinate Regression Image-Goal Navigation World Models for Robotics World Action Models Efficient Inference for World Action Models

Open paper AskPDF arXiv

World models capture environment dynamics and support counterfactual reasoning for planning. In image-goal navigation, where the target often lies beyond the current field of view, the agent must not only model environmental dynamics but also maintain a persistent spatial memory of the scene. Existing approaches either introduce spatial context by fine-tuning pretrained models on the target task, which underexploits their spatial priors, or rely on external 3D reconstruction, which is computationally expensive for real-world deployment. In this paper, we propose the Geometry Grounded World Action Model (G

^2

WAM) for image-goal navigation. G

^2

WAM encodes spatial cues through a learned scene-coordinate localization module that anchors the current, goal, and history frames in a common reference frame, yielding a stable and consistent spatial context that enables efficient navigation without explicit 3D reconstruction. Following the world-action-model paradigm, G

^2

WAM jointly predicts future states and actions, eliminating the need for costly cross-entropy method (CEM) style planning. In addition, G

^2

WAM operates entirely in latent space, avoiding the computational burden of pixel-level reconstruction and substantially reducing inference cost. Across four trajectory prediction benchmarks, G

^2

WAM reduces average ATE and RPE by 40.1% and 35.2%, respectively. In closed-loop navigation experiments, it improves the success rate from 54.0% to 63.0% and the success weighted by path length (SPL) from 36.4% to 59.2%, while maintaining high efficiency.

Abstract

World Models for Robotics

Explore similar work

G2^22WAM: World Action Models for Efficient and Geometry Grounded Image-Goal Navigation

SkipWorld: Latent Skip Planning with Distilled World Models for Efficient Visual Navigation

NavWAM: A Navigation World Action Model for Goal-Conditioned Visual Navigation

G $^2$ WAM: World Action Models for Efficient and Geometry Grounded Image-Goal Navigation