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

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Embodied AI is key to artificial general intelligence (AGI), requiring control of physical agents.
• Large language models (LLMs) and vision-language models (VLMs) have advanced AI capabilities.
• Vision-language-action (VLA) models bridge language understanding with physical action generation for embodied AI tasks.

Purpose of the Study:

• To provide the first comprehensive survey of vision-language-action (VLA) models in embodied AI.
• To establish a taxonomy for the rapidly evolving field of VLA models.
• To consolidate resources and identify future research trajectories.

Main Methods:

• Categorization of VLA research into three main areas: individual components, VLA-based control policies for low-level actions, and high-level task planners.
• Compilation of datasets, simulators, and benchmarks relevant to VLA research.
• Analysis of current challenges and future directions in embodied AI.

Main Results:

• A structured taxonomy detailing the landscape of VLA models.
• Identification of VLA applications in predicting actions and decomposing complex tasks.
• A comprehensive overview of essential resources for VLA research and development.

Conclusions:

• VLAs are a critical advancement for embodied AI, enabling robots to perform complex, language-conditioned tasks.
• The survey provides a foundational understanding and roadmap for future research in VLA models and embodied AI.
• Continued development in VLA models promises significant progress towards artificial general intelligence.