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.
References
This limitation points toward the ultimate goal and grand challenge for the field: the unsupervised discovery of the true underlying generative factors of an environment. Recent advances in self-supervised learning, vector quantization, and learned compression have made progress toward this goal, but the challenge remains open.
The method assumes domain features $\mathbf{x}_{\mathrm{dom}$ are identifiable---either explicitly (C-MAPSS) or via residual-based clustering (Scania APS); fully unsupervised invariance discovery is left to future work.
The coefficient read-outs reported here therefore condition on the reconstructed state, and closing the gap between supervised probes and unsupervised latent recovery remains open.
The clearest limitation is scope: our benchmarks are simulated or rendered, and DSReg has not yet been tested on a physical robot. That gap is the future work that excites us most. How does a robot behave when each factor of its world model is its own variable, and how far does that carry planning?