Sparse representation learning with modified q-VAE towards minimal realization of world model

Extraction of low-dimensional latent space from high-dimensional observation data is essential to construct a real-time robot controller with world model on the extracted space. However, there no established method for tuning dimension size automatically, suffering finding necessary and sufficient size, i.e. minimal realization model. In this study, we analyze improve Tsallis-based variational autoencoder (q-VAE), reveal that, under an appropriate configuration, it always facilitates making sparse. Even if pre-specified redundant compared realization, sparsification collapses unnecessary dimensions, allowing easy removal them. We experimentally verified benefits by proposed that can easily find six dimensions reaching task mobile manipulator requires six-dimensional state Moreover, planning such minimal-realization learned in was able exert more optimal action sequence real-time, reducing accomplishment time around 20%.