Learning can be scaled up across robot types

Robots make excellent specialists but are poor generalists. You have to create a different model for every task, robot and environment. It is often necessary to start from scratch when changing a single variable. What if we combined the knowledge from robotics to create a method of training a general-purpose, multi-purpose robot?

We are launching today a set of resources to support general-purpose robot learning, across different types or embodiments. We have combined data from 22 robot types with 33 partners from academic labs to create the Open X-Embodiment Dataset. We have also released RT-1X, a robots transformer (RT), derived from RT-1, and trained on our dataset. This model shows skills transfer across multiple robot embodiments.

We show that training a model using data from several embodiments results in significantly better performance than when trained with data from a single embodiment. Our RT-1X model was tested in five research labs and showed a 50% improvement on average for five commonly used robots compared with methods developed separately and specifically for each robot. Our research also demonstrated that training the visual language action model RT-2 on data from different embodiments tripled their performance in real-world robotics skills.

来源和详细信息：
https://www.deepmind.com/blog/scaling-up-learning-across-many-different-robot-types