Concept动作模型
Action Models
Models that map perceptual input directly to robot control signals, typically represented by vision-language-action (VLA) models, which use a pretrained multimodal large model as a backbone and directly output joint positions or discrete action tokens. Their advantage is strong end-to-end generalization, but they lack explicit future prediction capability.
Timeline (last 90 days)
Oct 13
World Action Models fuse two technical lineages, World Models and Action Models: the former lets AI internally simulate environment dynamics and predict the consequences of actions, while the latter directly generates robot control signals from perceptual input.
Unverified50%