Pure Vision Language Action (VLA) Models: A Comprehensive Survey

Authors: Dapeng Zhang, Jing Sun, Chenghui Hu, Xiaoyan Wu, Zhenlong Yuan, Rui Zhou, Fei Shen, Qingguo Zhou

Submitted: September 23, 2025 (v1); revised November 10, 2025 (v3) | arXiv:2509.19012 [cs.RO, cs.AI]

Abstract

VLA models represent a paradigm shift from traditional policy-based control to generalized robotics, transforming vision-language models into active agents for robotic manipulation and decision-making in complex environments. The survey systematizes over 300 recent studies.

Five VLA Paradigms (Taxonomy)

1. Autoregression-based — token prediction over action sequences (e.g., RT-2, OpenVLA)
2. Diffusion-based — denoising action distributions (e.g., Diffusion Policy, π0)
3. Reinforcement-based — RL fine-tuning of VLA policies
4. Hybrid — combinations of the above
5. Specialized — task-specific or embodiment-specific VLAs

Coverage

• VLA methodology taxonomy and systematic analysis
• Foundational datasets, benchmarks, and simulation platforms
• Applications across manipulation, navigation, and multi-task scenarios
• Challenges and future research directions toward scalable, general-purpose VLA systems