Hybrid Representation Learning for Robotic Manipulation

A new pretraining framework for robotic manipulation learns hybrid structural latent points by inserting a point-wise latent variational autoencoder into a point-cloud autoencoder's latent space. This approach jointly regularizes point-wise features and coordinates toward a Gaussian prior, producing a compact latent that captures coarse structural tendencies without precise geometry. The method combines the expressiveness of implicit neural fields with explicit geometric cues, addressing limitations of both fully implicit and fully explicit representations in 3D-aware pretraining. The framework aims to improve visual representations for embodied perception and manipulation tasks.