面向具身智能的视觉—语言—动作模型研究进展

Key Points

• The aim is to explore the development and progress of vision-language-action models in embodied intelligence.
• Reviewed the evolution of visual and language foundational models.
• Analyzed key technical modules of VLA including visual encoding and language representation.
• Classified existing models into single system, dual system, and hierarchical categories.
• Summarized pre-training and post-training strategies.
• Reviewed evaluation benchmarks in simulation and real-world environments.
• Identified challenges in real-time reasoning efficiency and data quality.
• Discussed issues related to environment generalization and safety ethics.
• Provided a comprehensive reference for future VLA research directions.

Abstract