This repository serves as a comprehensive survey of "World Models for Embodied AI," offering a curated collection of research papers and projects. It aims to provide a structured overview of advancements in creating internal simulators for environmental dynamics, enabling agents to perform forward and counterfactual rollouts for perception, prediction, and control across diverse tasks and domains. The collection is valuable for researchers and practitioners seeking to understand the state-of-the-art in this rapidly evolving field.

How It Works

World models, as presented in this survey, function as internal simulators that learn to predict future states of an environment based on current observations and actions. This allows embodied AI agents to "imagine" or "roll out" potential future scenarios, facilitating planning, decision-making, and control in complex, dynamic environments. The surveyed approaches encompass various architectural choices, including latent vector, sequential, and global representations, often leveraging deep learning techniques like transformers, diffusion models, and state-space models to capture intricate environmental dynamics.

Quick Start & Requirements

This repository is a curated list of research papers and projects, not a single runnable codebase. Therefore, there are no direct installation or execution instructions. Users interested in specific world model implementations must refer to the individual papers and their associated code repositories, which are linked within the survey.

Highlighted Details

• Broad Application Spectrum: The surveyed world models are applied across a wide range of embodied AI domains, including autonomous driving, robotic manipulation, navigation, and video generation.
• Chronological and Categorical Organization: Papers are organized by year and further categorized by their underlying representation and sequential/global processing approach (e.g., Latent Vector, Spatial Latent Grid, Decomposed Rendering Representation), providing a clear historical and technical progression.
• Diverse Methodologies: The collection showcases a variety of cutting-edge techniques, from transformer-based models and diffusion models to state-space models and Gaussian representations, highlighting the diverse research landscape.

Maintenance & Community

Information regarding project maintainers, community channels (e.g., Discord, Slack), or ongoing development status is not present in the provided survey details. It appears to be a static collection of research resources rather than an actively maintained software project.

Licensing & Compatibility

No specific software license is mentioned for the survey itself or for the aggregated projects. Users must consult the individual licenses of any linked research codebases for compatibility and usage restrictions.

Limitations & Caveats

This resource is a survey and collection of links to external research papers and their associated code. It does not provide a unified framework or a single point of entry for implementing or experimenting with world models. Users must independently locate, download, and integrate the code for each specific project of interest, potentially facing varying setup requirements, dependencies, and licenses for each.