LDA-1B: A 1B Robot Foundation Model Trained on 30k Hours of Embodied Data

chris_j_paxton · x · 2026-08-13

Researchers introduced LDA-1B, a dynamics-centric robot foundation model. Trained on EI-30k, a dataset comprising over 30,000 hours of heterogeneous embodied data, the model learns from diverse human demonstrations and dexterous manipulation tasks.

The core innovation of LDA-1B lies in its unified multimodal diffusion transformer framework, which jointly learns forward/inverse dynamics, visual forecasting, and policy within a single DINO latent space. This approach avoids redundant pixel-space appearance modeling, allowing the model to focus on task-relevant dynamics features and overcome the scaling limitations of traditional Behavior Cloning (BC).

Related event: Peking University and Partners Unveil Embodied AI Model LDA-1B(2 posts)→

Original post →

More from Embodied

Embodied channel →