LDA-1B: Unified World Model Trained on 30k Hours of Robot Data

chris_j_paxton · x · 2026-08-24

LDA-1B is a unified foundation model trained on 30,000 hours of human and robot interaction data. Instead of choosing between a world model and a language-conditioned policy, LDA-1B jointly learns forward dynamics, action prediction, and visual forecasting within a structured DINO latent space. This avoids pitfalls of pixel-level prediction, works on both dexterous hands and simple grippers, and generalizes across objects and tasks.

Related event: LDA-1B: unified robot world model trained on 30k hours(4 posts)→

Original post →

More from Embodied

Embodied channel →