World Dreamer v5-lite: A Multi-Robot World Action Model with One Dreamer

Guoning Wu; Zijian Cai; Yuhang Zhang; Ying Liu; Yongbin Zheng

cs.ROOct 27, 2026

World Dreamer v5-lite: A Multi-Robot World Action Model with One Dreamer

Authors: Guoning Wu, Zijian Cai, Yuhang Zhang, Ying Liu, Yongbin Zheng

Organizations: World Agents

Abstract

Multi-robot control requires integrating distributed camera views while predicting separate action sequences. Most vision-language-action and world-action models target one robot. Combining multiple views alone does not establish spatial correspondence between them. World Dreamer v5-lite is a multi-robot world-action model built around a shared world state for robot-specific action prediction. Calibrated features from a pretrained video encoder populate this state in workspace coordinates; its spatial size remains fixed as views are added. A vision-language model provides scene context and one query per robot. The queries read the shared state and exchange pose-aware information, while their action streams remain separate. Masked-view cooperative training withholds selected camera views from both visual branches while retaining all action targets. This encourages the use of observations shared by other robots. Pretraining uses more than 10,000 hours of trajectories from 20 simulated robot types, RoboFactory, and real robots. On RoboFactory, World Dreamer v5-lite achieves 74.0% mean success, 23.7 percentage points above the strongest evaluated baseline. Across four real-robot tasks, it reaches 78.8% mean success, a margin of 13.8 points. It also reaches 58.08% overall on RoboCasa365.

Figures & tables

Figure 1: Overall architecture of World Dreamer v5-lite. Multi-robot images and language tokens enter the vision–language model (VLM) before the learnable agent queries, yielding a scene-prefix key/value (KV) cache and one hidden state per robot query. The video strip depicts a temporal window: observed video latents are encoded and combined with camera geometry to build the shared world state $\boldsymbol{\phi}$ ; future latents supply training supervision only. The Action Decoder uses the VLM outputs, $\boldsymbol{\phi}$ , proprioception, and multi-robot noisy actions to predict a flow velocity for each robot.

Figure 2: Constructing the shared world state $\boldsymbol{\phi}$ . Observed video tokens from available views are placed in 3D using the intrinsics $K$ , the extrinsics, and optional depth. Each token is fused with its geometric embedding ( $+$ ) and written to the XY, XZ, and YZ planes. The right-hand panels schematically show how $\boldsymbol{\phi}$ is updated over successive observed frames.

Figure 3: Action Decoder information flow. Pink denotes agent queries and blue denotes noisy action tokens. From bottom to top, the Field Injection Block reads $\boldsymbol{\phi}$ into each query, the Scene Decoder Block conditions the causal action stream of each robot on the VLM key/value (KV) cache, and the Agent Interaction Block exchanges queries through pose-aware attention. Right: the relative poses $\mathbf{P}_{ij}$ bias inter-robot attention. The action streams remain separate across robots.

Method	Place Food	Camera Alignment	Take Photo	Mean
Diffusion Policy [ 6 ]	12	3	0	5.0
Global GauDP [ 23 ]	18	28	5	17.0
Local GauDP [ 23 ]	16	17	3	12.0
$\pi_{0.5}$ [ 2 ]	58	67	26	50.3
ACT [ 28 ]	13	21	10	14.7
Fast-WAM [ 25 ]	23	56	16	31.7

Table 1: Success rates (%) on RoboFactory. All methods use the same camera configuration; Mean is the unweighted average over the three tasks.

Figure 4: Success rates (%) on the 50 RoboCasa365 evaluation tasks, reported as the overall score and for the three splits. Baseline results are taken from the official leaderboard [ 19 ] ; the overall score is the average over all 50 tasks.

Method	Block transfer	Cap unscrewing	Starfruit bagging	Block passing	Mean
Diffusion Policy [ 6 ]	31	18	16	26	22.8
Global GauDP [ 23 ]	44	25	29	33	32.8
Local GauDP [ 23 ]	40	21	18	22	25.3
$\pi_{0.5}$ [ 2 ]	77	64	58	61	65.0
ACT [ 28 ]	58	33	41	53	46.3
Fast-WAM [ 25 ]	71	51	63	69	63.5

Table 2: Success rates (%) on the real-robot tasks. Mean is the unweighted average over the four tasks.

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 384 papers

Motubrain: An Advanced World Action Model for Robot Control
Apr 30, 2026Motubrain Team, Chendong Xiang, Fan Bao +17Robotic Control World Models for Robotics
Motubrain Team
UniviewVLA: A Unified Multiview Vision-Language-Action Model with World Modeling
Jun 19, 2026Tao Xu, Runhao Zhang, Zhijian Huang +7World Models for Robotics Multi-View Learning
1Tongji University · 2Shanghai Innovation Institute · 4Xiaomi EV +1
DuoMind: Enabling Distributed Multi-Robot Coordination with Semantic Communication
Oct 1, 2026Hanchu Zhou, Dechen Gao, Hang Wang +5Semantic Communication Multi-Robot Coordination
University of California, Davis · Microsoft Research · Analog Devices

Apr 30, 2026J/K move · Enter open · S save

Motubrain: An Advanced World Action Model for Robot Control

Motubrain Team, Chendong Xiang, Fan Bao, Haitian Liu+16

Motubrain Team

Robotic Control World Models for Robotics Action-Conditioned World Models Efficient VLA Model Inference Cross-Embodiment Robot Learning Efficient VLA Models+4

Open paper AskPDF arXiv

Vision-Language-Action (VLA) models generalize semantically well but often lack fine-grained modeling of world dynamics. We present Motubrain, a unified World Action Model that jointly models video and action under a UniDiffuser formulation with a three-stream Mixture-of-Transformers architecture. A single model supports policy learning, world modeling, video generation, inverse dynamics, and joint video-action prediction, while scaling to heterogeneous multimodal data such as video-only, task-agnostic, and cross-embodiment robot data. Building on Motus, Motubrain further introduces unified multiview modeling, an independent text stream for stronger language-action coupling, a shared cross-embodiment action representation, and an efficient post-training and deployment recipe for long-horizon real-world control. Our inference stack combines step reduction, compilation, FP8 quantization, DiT caching, V2A-style action-only inference, and real-time chunked closed-loop execution, achieving over 50x speedup over a naive baseline and up to 11 Hz inference. Experimentally, Motubrain achieves 95.8% and 96.1% average success on RoboTwin 2.0 under clean and randomized settings, respectively, attains the strongest reported EWMScore in our WorldArena comparison, and adapts to new humanoid embodiments with only 50--100 trajectories. These results show that unified world action models can scale in generality, predictive accuracy, and real-world deployability.

Abstract

World Models for Robotics

Explore similar work

Motubrain: An Advanced World Action Model for Robot Control

UniviewVLA: A Unified Multiview Vision-Language-Action Model with World Modeling

DuoMind: Enabling Distributed Multi-Robot Coordination with Semantic Communication