Learning High-Frequency Continuous Action Chunks in Latent Space

Researchers improved high-frequency robot control by learning action chunks in latent space with a variational autoencoder.

A reuse-then-refine strategy helps reduce discontinuities between chunks during real-time execution.

The approach improves smoothness, temporal consistency, and spatial consistency in continuous robot motion.

It is especially useful for contact-rich tasks that require precise, reliable control at very high rates.