LAPT-Scale introduces a scalable latent‑action pre‑training framework that leverages large‑scale egocentric human videos to improve generalist robotic policies. The system consists of two modules: LAPT-Scale-LAM learns continuous latent actions by explaining frame‑to‑frame visual transitions in human videos and uses these as human‑derived supervision; LAPT-Scale-Policy builds a VLA (vision‑language‑action) policy through latent‑action pre‑training, aligns it with the target robot embodiment, and then fine‑tunes on specific tasks.
Unlike prior vector‑quantized LAMs, LAPT-Scale-LAM adopts a continuous latent space, preserving fine‑grained motion and interaction dynamics. To go beyond pixel‑level reconstruction, the framework combines perceptual, semantic, and optical‑flow alignment signals, providing richer transition‑level supervision.
The model is pre‑trained on 23.3K hours of egocentric video, and the implicit action supervision yields consistent scaling benefits. When the annotated human video volume grows from 1K to 5K hours, real‑robot success rates improve from 63.3% to 85.0%. The learned policy also generalizes well to unseen visual conditions.
Review: LAPT-Scale demonstrates that massive human video corpora can effectively bootstrap robot policy learning, with continuous latent representations and multimodal alignment being pivotal for cross‑domain transfer.