Open X-Embodiment (OXE)

Largest open robotics dataset: 1M+ episodes from 22 embodiments

License

CC-BY-4.0

Overview

Largest open-source robotics dataset with 1M+ robot episodes from 22 robot embodiments, collected across 34 research labs worldwide. Contains 60 datasets spanning 527 skills for training generalist robot policies.

Open X-Embodiment is the largest publicly available robotics dataset, aggregating over one million real robot manipulation episodes collected across 22 different robot embodiments from 34 research institutions worldwide. The dataset encompasses 60 distinct datasets spanning 527 different manipulation skills, providing unprecedented diversity in robot morphologies, tasks, and operating conditions. This massive collection enables cross-embodiment learning research, where policies trained on data from multiple robot types can generalize to new platforms and tasks. The standardized data format facilitates large-scale pretraining of foundation models for robotics, supporting research into general-purpose manipulation systems. By combining contributions from dozens of labs into a unified resource, Open X-Embodiment addresses the data scarcity challenge that has historically limited robotics machine learning research.

Why it matters

Open X-Embodiment represents the largest collaborative effort in robotics data collection, uniting 34 research labs to create a dataset that enables training general-purpose robot policies. It proved that cross-embodiment training can improve performance over single-dataset policies.

Problem it solves

Individual robotics labs collect small datasets on specific robots, making it impossible to train general-purpose policies. OXE aggregates diverse manipulation data at unprecedented scale, proving that cross-embodiment training transfers knowledge between different robot morphologies.

Key takeaways

• Over 1 million real robot episodes from 22 different robot embodiments
• Collected across 34 research institutions worldwide spanning 527 skills
• RT-X models trained on OXE outperform single-dataset specialists
• Standardized data format (RLDS) enables easy integration of new datasets
• Foundation for Google DeepMind RT-2-X generalist policy

Who it is for

• Robotics researchers training foundation models
• Labs building generalist manipulation policies
• Teams developing VLA (Vision-Language-Action) models
• Companies scaling robot learning beyond single-task demos

Limitations

• Heavily weighted toward tabletop manipulation tasks
• Quality and labeling consistency varies across contributing labs
• Requires significant compute to train on the full dataset
• Real-world transfer still requires domain adaptation per robot