Transition Model
A core component of world models responsible for predicting environmental state transitions. Given the current state and an action as inputs, it outputs a prediction of the next state, forming the foundational technology that enables an agent's predictive capability.
Timeline (last 90 days)
Dreamer V1's world model consists of four parts: the representation model, the transition model, the reward prediction head, and the observation prediction head, all built on RSSM, corresponding respectively to posterior inference, prior prediction, and supervision signal generation.
Prediction capability is typically implemented by learning a transition model: given the current state and action, the model predicts the next state.
All Facts (2)
Dreamer V1's world model consists of four parts: the representation model, the transition model, the reward prediction head, and the observation prediction head, all built on RSSM, corresponding respectively to posterior inference, prior prediction, and supervision signal generation.
50%UnverifiedPrediction capability is typically implemented by learning a transition model: given the current state and action, the model predicts the next state.
50%