Motus: A Unified Latent Action World Model

arXiv 25.12 2025 57.6 method

TLDR

Motus unifies understanding, video generation, and action into a single latent action world model using MoT architecture and optical flow, achieving superior performance in simulation and real-world tasks.

Reasoning

The paper presents a novel unified architecture that integrates multiple modalities and modeling modes, with strong empirical results in both simulated and real-world settings. However, the abstract lacks detailed ablation studies and clarity on the limitations of the approach.

Read-first score

Read-first score 57.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 51.

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Topical relevance 42%

72.9

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

Methodology quality 25%

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

Reproducibility 25%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 446.

Keyword Scores

world model

generative world model

video world model

interactive world model

world dynamics prediction

world simulator

model-based reinforcement learning world model

Deep Analysis

Innovations

Unified latent action world model that integrates understanding, video generation, and action into a single system
Mixture-of-Transformer (MoT) architecture with three experts (understanding, video generation, action)
UniDiffuser-style scheduler enabling flexible switching between multiple modeling modes (world models, VLA, inverse dynamics, video generation, video-action joint prediction)
Optical flow for learning latent actions, extracting pixel-level 'delta action'
Three-phase training pipeline and six-layer data pyramid for large-scale action pretraining

Methodology

Motus employs a Mixture-of-Transformer (MoT) architecture with three experts (understanding, video generation, action) and a UniDiffuser-style scheduler to flexibly switch between modeling modes. It leverages optical flow to learn latent actions and adopts a three-phase training pipeline with a six-layer data pyramid to extract pixel-level 'delta action' for large-scale action pretraining.

Key Results

Motus achieves superior performance against state-of-the-art methods in simulation (+15% over X-VLA, +45% over Pi0.5) and real-world scenarios (+11~48% improvement), demonstrating that unified modeling of all functionalities and priors significantly benefits downstream robotic tasks.

Resources

Paper