LingBot-Touch paper finds masked latent prediction beats MAE and VAE for tactile models

A LingBot-Touch paper from Robbyant reports that masked latent prediction gives the best results among the training objectives it tested for a tactile foundation model.

The model is a shared Transformer that takes tokens from heterogeneous tactile sensors, and the authors compared it against MAE and VAE objectives trained on the same data. Elvis Saravia, who reviewed the paper on X, describes it as part of the JEPA line of research.