MetaWorld-X: Hierarchical World Modeling via VLM-Orchestrated Experts for Humanoid Loco-Manipulation

Yutong Shen; Hangxu Liu; Penghui Liu; Jiashuo Luo; Yongkang Zhang; Rex Morvley; Chen Jiang; Jianwei Zhang

cs.ROMar 9, 2026

MetaWorld-X: Hierarchical World Modeling via VLM-Orchestrated Experts for Humanoid Loco-Manipulation

Authors: Yutong Shen, Hangxu Liu, Penghui Liu, Jiashuo Luo, Yongkang Zhang, Rex Morvley, Chen Jiang, Jianwei Zhang,

Organizations: School of Information Science and Technology, Beijing University of Technology, Beijing, China · School of Information Science and Engineering, Fudan University, Shanghai, China · University of Alberta, Canada · University of Hamburg, Hamburg, Germany

Abstract

Learning natural, stable, and compositionally generalizable whole-body control policies for humanoid robots performing simultaneous locomotion and manipulation (loco-manipulation) remains a fundamental challenge in robotics. Existing reinforcement learning approaches typically rely on a single monolithic policy to acquire multiple skills, which often leads to cross-skill gradient interference and motion pattern conflicts in high-degree-of-freedom systems. As a result, generated behaviors frequently exhibit unnatural movements, limited stability, and poor generalization to complex task compositions. To address these limitations, we propose MetaWorld-X, a hierarchical world model framework for humanoid control. Guided by a divide-and-conquer principle, our method decomposes complex control problems into a set of specialized expert policies (Specialized Expert Policies, SEP). Each expert is trained under human motion priors through imitation-constrained reinforcement learning, introducing biomechanically consistent inductive biases that ensure natural and physically plausible motion generation. Building upon this foundation, we further develop an Intelligent Routing Mechanism (IRM) supervised by a Vision-Language Model (VLM), enabling semantic-driven expert composition. The VLM-guided router dynamically integrates expert policies according to high-level task semantics, facilitating compositional generalization and adaptive execution in multi-stage loco-manipulation tasks.

Figures & tables

Fig. 2 : MetaWorld-X achieves natural humanoid control through the dynamic orchestration of expert policies guided by a Vision-Language Model (VLM). The framework consists of two core modules: the Skill Expert Pool (SEP) trains specialized policy networks for fundamental motor skills using human motion data, while the Intelligent Router Module (IRM) employs a VLM as a supervisory teacher. The VLM guides the router’s training via few-shot inference, enabling a seamless transition from supervised learning to autonomous operation.

Fig. 3 : Architecture of the SEP module. We project human motion priors to the robot’s configuration space via operator $\mathcal{M}$ , then utilize alignment operator $\mathcal{A}$ to compute tracking signals for expert policies $\pi_{\theta_{i}}$ . By optimizing $\mathcal{J}_{\text{SEP}}$ with a dynamic weighting mechanism that prioritizes poorly-tracked joints, the module yields the converged expert policies.

Fig. 4 : Example of VLM prompt for Semantic Router.

Algorithm	Stand		Walk		Run		Sit		Carry
Algorithm	R. $\pm$ stderr.	S.(M)	R. $\pm$ stderr.	S.(M)	R. $\pm$ stderr.	S.(M)	R. $\pm$ stderr.	S.(M)	R. $\pm$ stderr.	S.(M)
PPO [ 28 ]	28.8 $\pm$ 0.4	5.3	28.8 $\pm$ 0.4	5.3	28.8 $\pm$ 0.4	5.3	28.8 $\pm$ 0.4	5.3	25.9 $\pm$ 0.4	5.4
SAC [ 29 ]	93.6 $\pm$ 17.7	4.9	32.7 $\pm$ 11.4	5.4	18.3 $\pm$ 1.5	5.4	275.9 $\pm$ 15.7	6.0	43.4 $\pm$ 9.7	5.3
TD-MPC2 [ 7 ]	749.8 $\pm$ 63.1	1.8	644.2 $\pm$ 162.3	1.8	66.1 $\pm$ 4.7	2.0	733.9 $\pm$ 120.6	1.1	438.0 $\pm$ 72.9	1.9
DreamerV3 [ 8 ]	699.3 $\pm$ 62.7	5.5	428.2 $\pm$ 14.5	6.0	298.5 $\pm$ 84.5	6.0	709.6 $\pm$ 98.4	5.2	427.8 $\pm$ 51.2	6.0
Ours (IRM)	815.9 $\pm$ 0.3 $\uparrow$	0.6 $\downarrow$	1118.7 $\pm$ 7.1 $\uparrow$	0.5 $\downarrow$	2056.9 $\pm$ 13.6 $\uparrow$	1.0 $\downarrow$	862.2 $\pm$ 2.1 $\uparrow$	0.6 $\downarrow$	963.5 $\pm$ 5.1 $\uparrow$	0.5 $\downarrow$

TABLE I: Expert policy evaluation: Our IRM module vs. baselines. We report maximum returns (

R.\pm stderr.

) and convergence steps

S

in Millions (M).

\uparrow

indicates higher is better,

\downarrow

indicates lower is better.

Fig. 5 : We evaluate the performance of four foundational motor skills: Walking, Running, Standing, and Sitting , after 50w training steps. The dashed lines in the learning curves qualitatively represent the reward thresholds associated with the successful execution of each skill.

Fig. 6 : In comparison with TD-MPC2, Green shaded areas highlight the superior performance of Ours , whereas red regions indicate limitations observed in baseline algorithms.

Method	Stand	Walk	Run	Sit	Carry
PPO [ 28 ]	0/10	0/10	0/10	0/10	0/10
SAC [ 29 ]	1/10	1/10	0/10	2/10	1/10
DreamerV3 [ 8 ]	2/10	2/10	1/10	3/10	2/10
TD-MPC2 [ 7 ]	3/10	3/10	2/10	4/10	3/10
Ours	9/10	9/10	9/10	8/10	9/10

TABLE II: Performance comparison on locomotion tasks (Success Rate / 10). Ours outperforms all baseline RL and model-based algorithms.

Fig. 7 : We evaluated the success rates of our method and Others on basic expert skills over 500k training iterations. For each task, ten independent trials were conducted. A trial was considered successful if the agent did not fall within a 30‑second episode and complete the task. The results demonstrate that after 500k steps, our method achieves a significantly higher success rate than Others. This confirms the robustness and reliability of our expert‑policy learning in these locomotion tasks.

Fig. 8 : The effectiveness of each ablated method on the door-opening task is illustrated in this figure.

Method	Suceess	steps(w)	return
TD-MPC2 [ 7 ]	$\checkmark$	32.38	198.42
w/o Router	$\checkmark$	20.36	296.57
w/o VLM	$\times$	NULL	NULL
w/o IL	$\times$	$\infty$	193.61
Full Model	$\checkmark$	12.64	303.95

TABLE III : Among the ablated variants, the model without Imitation Learning (IL) exhibits poor performance, primarily due to its incompatibility with our subsequent training and adaptation pipeline. Meanwhile, the without Vision-Language Model (VLM) configuration is not trained, as we aim to directly employ a rule-based router to achieve zero-shot compositional generalization without semantic guidance.

Fig. 9 : This figure illustrates a performance comparison between Ours and several baselines across various manipulation tasks. The plot includes a ”Reward Switcher” curve, which indicates the specific reference expert policy being utilized for reinforcement learning at each stage of the training process.

Task	Router (M)	PPO [ 28 ]	SAC [ 29 ]	TD-MPC2 [ 7 ]	DreamerV3 [ 8 ]	Ours
Door	0.17(S) $\to$ 0.74(W) $\to$ 2.0(R)	$58.0\pm 13.5$	$168.7\pm 3.9$	$285.0\pm 12.0$	$257.7\pm 37.9$	$\mathbf{470.0\pm 2.2}$
Basketball	0.87(S) $\to$ 2.0(R)	$58.0\pm 1.3$	$45.4\pm 5.5$	$148.4\pm 3.3$	$77.6\pm 1.2$	$\mathbf{250.0\pm 11.9}$
Push	1.09(S) $\to$ 2.0(R)	$36.8\pm 11.0$	$-201.3\pm 13.6$	$-113.8\pm 6.8$	$-59.9\pm 7.3$	$\mathbf{70.0\pm 2.1}$
Truck	0.57(S) $\to$ 1.61(W) $\to$ 2.0(R)	$58.0\pm 13.5$	$1142.6\pm 15.8$	$1213.2\pm 1.1$	$1383.4\pm 2.5$	$\mathbf{1500.0\pm 15.6}$
Package	0.51(S) $\to$ 2.0(R)	$-8112.3\pm 191.1$	$-6662.1\pm 137.2$	$-6788.5\pm 552.7$	$-5817.2\pm 903.8$	$\mathbf{-5200.0\pm 47.2}$

TABLE IV : Performance comparison on manipulation tasks. Routing schedules are indicated in Million steps (M) under each task, where S , W , and R denote Stand , Walk , and Reach reference stages, respectively.

Humanoid Loco-Manipulation

Humanoid loco-manipulation studies how robots coordinate walking, balance, and whole-body interaction to move and handle objects.

See all 93 papers

LUCID: Latent-Skill Unified Control via Imagined Dynamics for Long-Horizon Humanoid Loco-Manipulation
Aug 7, 2026Cheng Guo, Mingzhe Ni, Angelo Cangelosi +1Humanoid Loco-Manipulation Latent Action Learning
Department of Computer Science, University of Manchester, Manchester, UK · Human-Robot Interfaces and Interaction Laboratory, Italian Institute of Technology, Genoa, Italy
HEX: Humanoid-Aligned Experts for Cross-Embodiment Whole-Body Manipulation
Apr 9, 2026Shuanghao Bai, Meng Li, Xinyuan Lv +15Flow Matching for Action Generation Robotic Control
Beijing Innovation Center of Humanoid Robotics · Xi’an Jiaotong University · Nankai University +1
X-Loco: Towards Generalist Humanoid Locomotion Control via Synergetic Policy Distillation
Mar 4, 2026Dewei Wang, Xinmiao Wang, Chenyun Zhang +4Humanoid Robot Locomotion Robot Policy Learning
University of Science and Technology of China · Institute of Artificial Intelligence (TeleAI), China Telecom · Harbin Engineering University

Aug 7, 2026J/K move · Enter open · S save

LUCID: Latent-Skill Unified Control via Imagined Dynamics for Long-Horizon Humanoid Loco-Manipulation

Cheng Guo, Mingzhe Ni, Angelo Cangelosi, Arash Ajoudani

Department of Computer Science, University of Manchester, Manchester, UK · Human-Robot Interfaces and Interaction Laboratory, Italian Institute of Technology, Genoa, Italy

Humanoid Loco-Manipulation Latent Action Learning Robot Skill Learning Model-Based RL Robotic Loco-Manipulation Loco-Manipulation

Open paper AskPDF arXiv

Long-horizon humanoid loco-manipulation requires composing versatile whole-body skills and reliable high-level decision making. Existing methods often coordinate pretrained skills with scripted planners, finite-state machines or task-specific model-free policies, restricting their ability to handle complex task sequences. To address this limitation, we propose LUCID, a hierarchical model-based reinforcement learning framework that plans over reusable skills through imagined rollouts of a learned dynamics model. LUCID first trains a structured latent-conditioned low-level policy via adversarial imitation and then freezes it while jointly learning a high-level policy and macro-dynamics world model. The world model predicts the temporally extended state transitions induced by latent decisions, enabling high-level policy optimization through imagined rollouts. We evaluate our framework across various simulated multi-object rearrangement scenarios. Experimental results show that LUCID improves the full-task success and partial-completion rates compared to prior baseline methods, demonstrating its effectiveness in complex sequential loco-manipulation tasks.

Abstract

Humanoid Loco-Manipulation

Explore similar work

LUCID: Latent-Skill Unified Control via Imagined Dynamics for Long-Horizon Humanoid Loco-Manipulation

HEX: Humanoid-Aligned Experts for Cross-Embodiment Whole-Body Manipulation

X-Loco: Towards Generalist Humanoid Locomotion Control via Synergetic Policy Distillation