A Comprehensive Overview of Vision-Language-Action Models

Vision-Language-Action (VLA) models combine computer vision and natural language processing to enable general-purpose robotics. Google DeepMind's RT-1 demonstrated transformers could map visual and linguistic inputs to robotic actions, while RT-2 incorporated pre-trained vision-language models for broader semantic understanding. The Open X-Embodiment dataset pooled data from 22 institutions across 20+ robot types, enabling better cross-platform generalization. OpenVLA is a 7B parameter open-source model trained on 970k episodes. Physical Intelligence's π-series represents ongoing efforts to build general-purpose foundation models for robotics, with capabilities ranging from dexterous manipulation to autonomous task completion.