The robot agent (ORA) runs each task on the same unseen layouts of the twin, once without the demonstration context and once with it. Each row is one layout: without the context on the left, with it on the right.
The videos are the twin's scene camera, sped up; each caption gives the outcome and the agent's own run time.