Unsupervised discovery of the true generative factors of environments

Develop an unsupervised representation learning method that discovers and disentangles the true underlying physical state variables (generative factors) of a deterministic environment from high-dimensional exteroceptive observations and actions, yielding a representation equivalent to the irreducible physical state.

Background

The authors emphasize that even with improved geometric regularization, learned latents may still deviate from the true physical states, and small residual misalignments can accumulate over very long horizons. They therefore highlight a broader, foundational objective for world models: to recover the environment’s true generative factors in an unsupervised manner so that the learned representation is equivalent to the irreducible physical state.

They note that recent progress in self-supervised learning, vector quantization, and learned compression has moved toward this goal, but the problem remains unresolved, and achieving such representations would enable robust, law-faithful cloning rather than merely plausible prediction.