Training one policy to handle multiple tasks via language conditioning. A task embedding tells the model what to do, while the shared vision-action backbone decides how. Builds on the Demo Collection & VLA lessons.