MotionWAM: Towards Foundation World Action Models for Real-Time Humanoid Loco-Manipulation

arXiv 2026 55.1 method, application

TLDR

MotionWAM enables real-time humanoid loco-manipulation by conditioning policy on video world model features, outperforming baselines on real-world tasks.

Reasoning

Strengths include real-time performance, real-world experiments on nine tasks, and significant improvement over VLA baselines. Weaknesses are not explicitly discussed in the abstract, but potential limitations in generalization or complexity are not addressed.

Read-first score

Read-first score 55.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 39.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Citation impact 18%

96.2

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.96179995

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,result

Topical relevance 29%

55.7

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FoundationFrontierBridge

Rank sensitivity

Stability: volatile; rank range: 463.

Keyword Scores

world model

video world model

world dynamics prediction

generative world model

model-based reinforcement learning world model

interactive world model

world simulator

Deep Analysis

Innovations

Conditioning the policy on intermediate denoising features of a video world model for real-time humanoid loco-manipulation
Replacing the upper-lower body split with a unified motion latent and predicting whole-body motion tokens covering locomotion, torso motion, height regulation, foot interaction, and hand manipulation in a single action space
A three-stage learning framework that progressively adapts the video world model to egocentric visual dynamics and to the target humanoid embodiment

Methodology

MotionWAM uses a three-stage learning framework to adapt a video world model to egocentric visual dynamics and the target humanoid embodiment. The policy is conditioned on intermediate denoising features of the video world model, and whole-body motion tokens are predicted in a unified action space, replacing the traditional hierarchical upper-lower body split.

Key Results

On nine real-world Unitree G1 tasks, MotionWAM runs in real time and substantially outperforms Vision-Language-Action (VLA) baselines fine-tuned on the same demonstrations by over 30% in overall success rate, while also executing task-driven foot interaction that decoupled upper-lower policies cannot achieve.

Resources

Paper