DeepMind unites 33 labs to build an ImageNet moment for robots

Google DeepMind, in collaboration with 33 academic labs across more than 20 institutions, has released Open X-Embodiment, a massive open-source robotics dataset paired with new general-purpose RT-X models. The goal is straightforward: instead of training a separate model from scratch for every unique robotic arm or gripper, you train one foundational model across diverse hardware embodiments.

In traditional robotics, changing a single variable like robot geometry usually breaks existing policies. This release provides a unified foundation to let robots learn from each other's experiences.

22 robot types and over a million episodes

Open X-Embodiment combines demonstrations across 22 different robot types, covering more than 500 skills and 150,000 tasks across over 1 million episodes. DeepMind likens it to an ImageNet moment for robotics, making it the most comprehensive dataset of its kind.

Using this pool of data, DeepMind trained two robotics transformer variants: RT-1-X (based on their real-world control model RT-1) and RT-2-X (based on their vision-language-action model RT-2, which co-trains on web data and robot actions).

Performance and emerging skills

Evaluated across five independent partner labs, RT-1-X improved average task success rates by 50% across five standard robots, beating specialized models trained solely on domain-specific data.

On the larger RT-2-X architecture, cross-embodiment training unlocked emergent generalization. When tested on skills and objects never seen in its own robot dataset but present in another robot's data, RT-2-X tripled the success rate compared to RT-2.

It also showed significantly improved spatial reasoning. Swapping an instruction from "move apple on cloth" to "move apple near cloth" yielded distinct, context-aware physical trajectories. DeepMind noted that such cross-platform skill transfer requires a sufficiently high-capacity architecture.

How to get it

The Open X-Embodiment dataset and the RT-1-X model checkpoints are now publicly available for the research community to download and build upon. Research teams interested in expanding the effort can also connect via the project's coordination group.

Why it matters

If you work on robotic learning, manipulation, or embodied AI, this is an essential resource to test. Collecting physical demonstration data is the single biggest bottleneck in robotics, and pooling cross-platform trajectories lowers the barrier for any lab without a massive hardware farm.

We think this is a smart shift toward treating physical actions like pre-trained language or vision features. The results prove that data diversity across different mechanical bodies actually helps individual robots perform better. However, RT-1-X is still primarily a research checkpoint; deploying these policies to custom setups in unstructured environments will still require domain fine-tuning and careful safety oversight.