Open X-Embodiment and RT-X: a shared robotics dataset from 22 robot types
In one sentence Google DeepMind and partners from 33 academic labs pool over one million trajectories from 22 robot types (527 skills) and train RT-1-X and RT-2-X, which improve over models trained on a single robot.
One of robotics' biggest problems is that each robot usually has its own dataset and its own model: what a robotic arm learns does not easily transfer to a different robot. Open X-Embodiment tries to break this barrier by pooling data from many robots.
Google DeepMind, with partners from 33 academic labs, merges 60 existing datasets into a single open dataset: over one million real trajectories from 22 robot types, from single arms to bi-manual robots and quadrupeds. With this data it trains RT-1-X and RT-2-X, versions of RT-1 and RT-2 that learn from all robots together.
Companies
Google DeepMind
Tools
RT-1-X, RT-2-X, Open X-Embodiment Dataset
Tags
Sources