# A Survey on Vision-Language-Action Models for Embodied AI

**Authors:** Yueen Ma, Zixing Song, Yuzheng Zhuang, Jianye Hao, Irwin King  
**arXiv:** [2405.14093](https://arxiv.org/abs/2405.14093)  
**Date:** 2024-05-23  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2405.14093v4/

## Key points

- Embodied AI is a crucial component of artificial general intelligence focusing on controlling physical agents in the real world
- Significant advancements have been made with vision-language-action models (VLAs) in recent years
- VLAs generate actions based on language inputs to address language-conditioned robotic tasks
- The survey by Yueen Ma, Zixing Song, Yuzheng Zhuang, Jianye Hao, and Irwin King categorizes VLAs into three main research lines: individual components, control policies for low-level action prediction, and high-level task planners
- This categorization enables VLAs to effectively follow general user instructions
- The survey highlights various VLAs developed recently and emphasizes the importance of systematic analysis for this rapidly evolving field
- Relevant resources such as datasets, simulators, and benchmarks are essential for advancing research in embodied AI with VLAs
- Challenges faced by VLAs in embodied AI are discussed along with promising future research directions
- A project associated with the survey is available at https://github.com/yueen-ma/Awesome-VLA

## Summary

Embodied AI is a crucial component of artificial general intelligence that focuses on controlling physical agents to carry out tasks in the real world. In recent years, significant advancements have been made in this field with the emergence of vision-language-action models (VLAs). These multimodal models leverage their ability to generate actions based on language inputs to address language-conditioned robotic tasks in embodied AI. A comprehensive survey conducted by authors Yueen Ma, Zixing Song, Yuzheng Zhuang, Jianye Hao, and Irwin King delves into the landscape of VLAs for embodied AI. The survey provides a detailed taxonomy of VLAs categorized into three main research lines: individual components of VLAs, control policies for low-level action prediction, and high-level task planners for breaking down complex tasks. This categorization enables VLAs to effectively follow general user instructions. The survey also highlights an array of VLAs that have been developed in recent years and emphasizes the importance of systematic analysis in capturing this rapidly evolving landscape. It offers insights into relevant resources such as datasets, simulators, and benchmarks essential for advancing research in this domain. Additionally, the authors discuss the challenges faced by VLAs in embodied AI and outline promising future directions for research and development. To further support researchers and practitioners interested in exploring and contributing to VLA advancements in embodied AI, the authors have created a project associated with this survey available at https://github.com/yueen-ma/Awesome-VLA. Overall, this survey serves as a valuable resource for those seeking to stay updated on the latest developments in vision-language-action models for embodied AI.