Comment by grumbelbart

Agreed. The solution will likely be some vision foundation model that directly sends controls to the robot ("move here, grab, move there"), trained by Amazon with RL to integrate collision avoidance, object detection, grasping point detection, grasp verification etc.