Open X-Embodiment and RT-X

Background

The project standardizes heterogeneous robot data into a shared format and studies RT-X models trained across many robot types and tasks. It tests whether diversity across embodiments improves generalization.

Problem

The work addresses a central constraint in Datasets: building systems that learn useful representations or actions while remaining general enough to transfer beyond a single demonstration or environment.

Core idea

Open X-Embodiment combines robot datasets across institutions and trains policies that transfer across embodiments.

Architecture and method

The project standardizes heterogeneous robot data into a shared format and studies RT-X models trained across many robot types and tasks. It tests whether diversity across embodiments improves generalization.

• Cross-institution dataset
• Standardized robot data
• Cross-embodiment policy training

Results and impact

It created a major open data foundation for studying robot learning at scale.

Prerequisites

• RT-1
• Imitation learning
• Robot datasets

Recommended reading order

Read the explanation above, review the related topic pages, then use the primary-source links below to inspect the abstract, figures, experiments, and released implementation.