MECo-WAM is proposed, a Multi-Expert Co-Training World Action Model that injects action-relevant 4D geometric priors into video-action representations while preserving the original lightweight inference graph.
Abstract
World Action Models (WAMs) have shown strong potential for robotic manipulation by jointly modeling visual future dynamics and executable action sequences. However, existing video-action co-training methods primarily optimize appearance-oriented video latents, which may insufficiently capture the temporally evolving geometry required for precise manipulation. We propose MECo-WAM, a Multi-Expert Co-Training World Action Model that injects action-relevant 4D geometric priors into video-action representations while preserving the original lightweight inference graph. During training, MECo-WAM combines video and action experts with a lightweight 4D expert supervised by relational targets from a frozen VGGT encoder. Asymmetric expert visibility prevents non-causal shortcuts from auxiliary geometry to action generation. To transfer geometric knowledge into the deployed video-action pathway, we introduce decayed 4D read-mask attention, which provides restricted current-frame geometric guidance early in training and progressively removes this dependency. We further propose action-aware temporal geometric distillation, which aligns within-frame geometric relations and their temporal evolution while emphasizing visual regions most relevant to robot actions. At deployment, all auxiliary 4D components are removed. Experiments on LIBERO (98.2%), RoboTwin 2.0 (92.6%), and challenging real-world manipulation tasks show that MECo-WAM improves manipulation performance without increasing inference cost.
World-Action Models (WAMs) jointly learn future visual prediction and action generation, using video dynamics as a representation-learning signal for robotic manipulation. However, their video latents are primarily optimized for visual prediction and are not explicitly encouraged to preserve cross-view geometric structure or spatially localized, object-relevant semantics. We propose \textbf{GaussianWAM}, a training-time representation-enhancement framework that organizes geometric and semantic supervision through a 3D Gaussian field. Given synchronized multi-view observations, frozen geometry and vision foundation models provide depth, camera parameters, and dense semantic features. GaussianWAM binds these heterogeneous signals to shared Gaussian primitives and renders spatially aligned semantic, depth, and coverage targets, which are distilled into the current-observation representations of the WAM. All teacher models, Gaussian components, and auxiliary prediction heads are removed after training, leaving the original WAM inference path without additional modules or forward computation. On LIBERO-Plus, GaussianWAM improves FastWAM from 52.05\% to 71.29\% and Cosmos Policy from 71.52\% to 77.30\%. Direct CLIP and VGGT distillation already establishes a strong FastWAM baseline of 69.37\%, while Gaussian-field unification further improves it to 71.29\%, supporting the benefit of spatially organizing heterogeneous teacher signals. GaussianWAM also improves performance on standard LIBERO and shows positive transfer trends on RoboTwin and real-world manipulation. These results suggest that training-time Gaussian distillation provides a practical way to inject geometry- and semantics-related supervision into WAM representations without changing their deployment architecture.
Zijian Zhang, Yuqing Jiang, Weitao Zhou et al.· 0 citations
World Action Models (WAMs) improve robot policy learning by jointly modeling future visual dynamics and actions. However, their scalability and generalization remain constrained by their reliance on costly expert demonstrations. We challenge this by asking whether future supervision for WAMs must originate from target-task expert trajectories. In this paper, we propose Vid2WAM, an offline distillation framework that transfers visual diffusion priors from a large video foundation model into a compact WAM student. Given an observation and language instruction, Vid2WAM distills supervision through two complementary channels: task-conditioned future rollouts directly supervise the student's future prediction branch, while an inverse dynamics model recovers embodiment-specific pseudo-actions for action learning. To robustly integrate synthetic and real supervision, we introduce source-aware residual action adaptation that learns source-specific corrections around a shared action backbone and mitigates interference from noisy pseudo-actions. During inference, both the video teacher and inverse dynamics model are discarded, leaving only the WAM student for efficient deployment. Simulation and real-world experiments demonstrate that Vid2WAM improves novel-task generalization and data efficiency under limited expert demonstrations while preserving low-latency inference.
Chenhao Qiu, Ruixiang Wang, Runyi Zhao et al.· 0 citations
LD4WAM is presented, which pairs a Latent Dynamics Model trained with semantic reconstruction and real motion alignment with a World Dynamics Action Model built as a mixture-of-transformers (MoT), which preserves full future-video generation and uses learnable queries to distill these latent dynamics from generated futures for action conditioning.
Zhen Shen, Jiaqi Liang, Jasper Lu et al.· 0 citations
World models enable agents to perform forward rollout and planning without real-world interaction. However, their application in open-world embodied intelligence remains limited by the high cost of action annotations and the heterogeneity of action spaces across platforms. Recently, latent action models (LAMs) have alleviated this bottleneck by learning action representations directly from unlabeled human videos in a self-supervised manner. Nevertheless, most existing LAMs rely on single-view inputs and operate primarily in 2D pixel space, raising a fundamental question: can simply incorporating multi-view videos into LAM training endow the learned latent actions with 3D-aware perception? Our study shows that the answer is negative. The primary reasons lie in future-frame appearance leakage as well as inter-camera appearance discrepancies and viewpoint variations. To address these issues, we propose LAWM-3D, which introduces three tightly coupled key designs: (1) a multi-view invariant unified action tokenization scheme for learning 3D-aware latent actions; (2) a geometric alignment constraint that anchors intermediate encoder features to a pretrained 3D foundation model, thereby explicitly providing cross-view geometric correspondences; and (3) a non-injective RGB-D joint reconstruction objective that prevents shortcut learning from future-frame appearance information, forcing the LAM to focus supervision on motion cues with geometric significance. Importantly, these components are not simply stacked but are tightly coupled through a unified motivation. Built upon a two-stage paradigm of large-scale human video pretraining followed by robot fine-tuning, extensive experiments demonstrate that the proposed 3D-aware latent actions significantly improve world model performance, achieving SOTA results in generation quality, physical consistency, and generalization ability.
Jiaru Yang, Jiale Zhange, Jiawei Li et al.· 0 citations
The Robust-WAM is a general post-training method for video-generation-based WAMs that preserves the VAE-based generative path and adds a lightweight semantic foresight alignment objective on the action stream to retain the large-scale VGM pretraining while grounding actions in appearance-invariant dynamics.
Haodong Yan, Junfeng Li, Junjie He et al.· 0 citations
Vision-Language-Action (VLA) policies have advanced language-conditioned robotic manipulation, yet action-imitation objectives provide only weak supervision for metric 3D structure and short-horizon physical evolution. Geometry-enhanced policies mainly improve current-scene grounding, whereas predictive policies often model future dynamics in RGB or latent spaces and may incur substantial deployment cost. GaussianDream demonstrates that training-time current Gaussian reconstruction and future Gaussian prediction provide effective 3D supervision, but its dense VGGT/TGE-based prefix jointly carries state, dynamics, and action-conditioning information. We present \textbf{\methodname}, a compact, policy-native extension that inserts \textbf{World State Tokens} and \textbf{World Prediction Tokens} directly into the VLA backbone. A training-only \textbf{World Representation Head} decodes these tokens into a Current World and coupled Future Prediction over shared Gaussian primitives, while static--dynamic factorization preserves persistent structure and focuses residual motion on interaction-relevant regions. At inference, the head, renderer, auxiliary objectives, and VGGT/TGE pathway are removed, leaving only 20 world tokens without online Gaussian decoding or rollout. \method achieves \textbf{98.6\%} on LIBERO and \textbf{87.8\%} on LIBERO-Plus, with clear gains under Camera and Layout shifts. Real-robot experiments further improve average success from 29.2\% to 52.5\% over reproduced $\pi_{0.5}$ while maintaining efficient closed-loop control.
Yuqing Jiang, Zijian Zhang, Weitao Zhou et al.· 0 citations
