LDA-1B: Scaling Latent Dynamics Action Model via Universal Embodied Data Ingestion

arXiv 2026 40.8 method, system

TLDR

LDA-1B scales robot foundation model by jointly learning dynamics, policy, and visual forecasting from diverse embodied data, outperforming prior methods.

Reasoning

The paper introduces a novel approach to scaling robot foundation models via universal data ingestion and latent dynamics prediction, with strong empirical results in simulation and real world. However, the reliance on a specific latent representation (DINO) and complex multi-modal architecture may limit generalizability.

Read-first score

Read-first score 40.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 42.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Topical relevance 29%

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=dataset,experiment

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=dataset

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 263.

Keyword Scores

world dynamics prediction

world model

interactive world model

generative world model

world simulator

video world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Universal embodied data ingestion by jointly learning dynamics, policy, and visual representations, assigning distinct roles to data of varying quality.
Scalable dynamics learning via prediction in a structured DINO latent space, avoiding pixel-space appearance modeling.
Multi-modal diffusion transformer to handle asynchronous vision and action streams, enabling stable training at the 1B-parameter scale.
Creation of EI-30k, a unified embodied interaction dataset with over 30k hours of human and robot trajectories.
Data-efficient fine-tuning that leverages low-quality trajectories (30% typically harmful) to gain 10% improvement.

Methodology

LDA-1B is a 1B-parameter robot foundation model that jointly learns dynamics, policy, and visual representations from a unified heterogeneous dataset (EI-30k). It uses a structured DINO latent space for dynamics prediction to avoid pixel-space modeling and a multi-modal diffusion transformer to process asynchronous vision and action streams.

Key Results

LDA-1B outperforms prior methods like π_{0.5} by up to 21%, 48%, and 23% on contact-rich, dexterous, and long-horizon tasks, respectively, and achieves data-efficient fine-tuning gaining 10% by leveraging 30% low-quality trajectories.

Resources

Paper