“Physical Intelligence's robotics model is a vision-language model adapted for motor control, featuring a vision encoder and an 'action expert' action decoder, structurally resembling a mixture-of-experts transformer.”