Can large, diverse pretrained models be used to consolidate robotic learning methods? In this paper, the authors explore the possibility of training a “generalist” X-robot policy that can adapt to new robots, tasks, and environments. They provide standardized datasets and models for robotic manipulation, showcasing the RT-X model trained on data from 22 different robots. The authors introduce the Open X-Embodiment Dataset, the largest open-source real robot dataset to date, containing over 1 million robot trajectories across 22 robot embodiments. They train two models, RT-1 and RT-2, on this data and demonstrate positive transfer and improved capabilities of multiple robots. RT-2-X outperforms RT-2 by 3x in emergent skill evaluations.
https://robotics-transformer-x.github.io/