Models & learning
Autoregressive model
An autoregressive model represents a sequence by predicting each later element from earlier elements in a fixed order. In robot learning, an autoregressive video model can learn how observed motion and interaction history leads to future visual states, but causal ordering alone does not guarantee accurate or useful control.
Also known as: AR model, autoregressive modelling, autoregressive modeling
Updated
Predicting a sequence in order
An autoregressive model factorises a sequence so each element is predicted from the elements before it. The element might be a text token, an image patch, a video block, or an action. At deployment, generated elements can become part of the context for later predictions.
This causal order differs from bidirectional masked modelling. A bidirectional model may use context on both sides of a missing element during pretraining. An autoregressive model cannot read the future element it is learning to predict.
Autoregressive video for robot learning
LongLive-2.0 applies block-causal autoregressive training to long video. Long-WAM continues that training on robot and egocentric video, then adapts the visual model to condition robot actions on observed history and predicted future latents.
The ordering matters because a deployed robot also sees observations arrive from past to present. Long-WAM reports larger benefits from longer visual history for autoregressively pretrained initializations than for a bidirectionally pretrained initialization adapted with the same causal policy mask.
Causality does not ensure reliability
Autoregressive generation can accumulate its own errors because later predictions depend on earlier generated states. Teacher forcing can make training efficient, but it also exposes the model to cleaner histories than it may produce at deployment. Techniques such as error recycling try to narrow that gap.
The label also says nothing about action safety, timing, embodiment fit, or physical accuracy. A robot system still needs aligned observations and actions, closed-loop evaluation, explicit latency measurements, and recovery behaviour.
Sources
Related terms
Models & learning
World model
A world model is a learned predictive model of how an environment changes, often conditioned on a robot’s current state and candidate actions. It may forecast pixels, latent states, rewards or other task-relevant quantities. A policy or planner can use those predictions to compare possible futures before commanding the physical robot.
Models & learning
World-action model
A world-action model is an emerging, non-standardised name for a robot model that couples prediction of future world states with generation of executable actions, often by adapting a video or world-model backbone as a policy. It differs from an action-conditioned world model that only predicts consequences of supplied actions, and from Barrett’s WAM, the Whole Arm Manipulator.
Models & learning
Observation horizon
An observation horizon is the span or number of recent sensor and state observations that a robot policy uses to predict its next action or action sequence. It is distinct from the action horizon, which describes how far into the future the policy's output extends.
Models & learning
Action chunking
Action chunking is a policy design in which one observation is used to predict a sequence of several future robot actions rather than a single next action. The chunk can capture short-term motion structure and reduce the effective number of sequential predictions, while its execution and replanning schedule determine how quickly the robot can react to new observations.
Models & learning
Policy
A policy is the decision rule that maps a robot’s current observations or estimated state, and sometimes a task instruction, to an action or probability distribution over actions. It can be hand-designed or learned from demonstrations, rewards or both. In humanoid robotics, its outputs may be joint targets, torques, end-effector changes or higher-level skills.