Adapting pre-trained visual features conditioned on the current task is a key design choice that enables a single multi-task policy to perform well across a wide variety of robotic manipulation and locomotion tasks.
