World model

What it means

In robot learning, a world model usually means a learned representation of how an environment evolves. It can summarize observations and predict future states or observations, often conditioned on actions. Ha and Schmidhuber's World Models separates a visual representation, a predictive memory model and a controller, making the prediction component distinct from the policy that chooses behavior. This entry uses that learned-model meaning. [1]

A model need not predict every pixel explicitly to be useful. Dreamer learns behavior through imagined sequences in a compact latent space. The predicted representation supports learning and decision-making even when it is not presented as a realistic video for a human viewer. [2]

How it applies to humanoids

For robots, predicting what may happen after a proposed action can support training or evaluation. 1X's world-model article shows multiple predicted futures from the same starting frames under different robot action commands. It discusses robot-policy evaluation as a motivation, including the difficulty of repeating experiments in changing physical environments. [3]

This is relevant to humanoid manipulation because action consequences involve the robot, objects and their contacts. However, the existence of a manufacturer's world-model research does not establish which model is installed on a particular product. Product-specific deployment needs its own source. [3]

Terminology and prediction traps

A world model and a policy are different components. The model predicts consequences, while the policy selects actions; an architecture may train or connect both. Dreamer uses a learned model to improve behavior, but calling a system a world model alone does not describe its entire controller. [1, 2]

Generated futures are predictions, not measurements of a completed physical task. 1X's examples show alternative action-conditioned outcomes, so a plausible generated video cannot establish that hardware successfully carried out the pictured sequence. Compare such predictions with real-world observations when assessing accuracy. [3]

A learned model can be exploited. World Models describes controllers discovering actions that work inside an imperfect model but fail in the actual environment, especially away from the training distribution. Visually convincing output therefore does not eliminate the need for model-error and transfer checks. [1]

A conventional physics simulator and a learned world model should also be identified separately. The cited projects learn predictive representations from experience; a model built from specified physical equations is a different modeling choice even if both can support planning. [1, 2]

How to read it on humanoidversus.com

Read ai_stack for explicitly named systems and follow the sources to determine whether a model is used for prediction, training, evaluation or deployed control. compute is hardware context, not a world-model accuracy score. A null stack means the record lacks a sourced name.

This entry has no live table or robot links. No product is associated with a world model solely because its manufacturer publishes model research. Benchmark results should retain their environment and metric, with predicted outcomes kept distinct from physical task results.

Further reading

World Models explains learned representations and model exploitation. Dreamer develops learning through latent imagination. 1X's article supplies a robot-specific account of action-conditioned prediction and evaluation. [1–3]