Robot Models
The models chapter covers the learned part of a robot: how vision-language-action models work, what separates one architecture from another, how they are trained and evaluated, and where they still fail.
§Articles in this chapter
1 articleVision-language-action models (VLAs), explained
A VLA model turns camera images and a text instruction into robot actions. How they work, how the main architectures differ, and where they fail.
§Terms from this chapter
Full glossary →Action chunking
Having a policy predict a short sequence of future actions at once instead of one action per step. It smooths motion and lets a slow model drive a fast control loop.
Imitation learning
Training a robot policy from demonstrations of the task, usually collected by a human teleoperating the robot, rather than from a reward signal.
Policy
The function that maps what a robot observes to what it does next. In learned systems the policy is a neural network; in classical systems it is hand-written control logic.
Reinforcement learning
Training a policy by trial and error against a reward signal. In robotics it is mostly done in simulation, where millions of attempts are cheap and failures break nothing.
Robot foundation model
A large model pretrained on broad robot and non-robot data that can be adapted to many tasks and robot bodies, instead of being trained from scratch for one.
Vision-language-action model
A model that takes camera images and a natural-language instruction as input and outputs robot actions. Most are built by adding action output to a pretrained vision-language model.
World model
A model that predicts how the environment will change in response to actions. Robots can use one to plan, to evaluate a policy without running it, or to generate synthetic training data.