TimWM studies how to represent the effect of an action separately from the current state of an environment.
The intended model encodes a sequence of observations, predicts an action-effect representation, and uses that representation to predict the next state. Auxiliary predictions can describe properties such as speed, heading, collision, or progress.
The project keeps those properties distinct from the main state representation. A policy can then use the learned state and selected predictions while the world model remains fixed.
The current work is a rebuild of an earlier prototype. The main questions concern the representation of action effects and whether it supports useful prediction and control. The design itself does not establish that benefit.