When a vision-based grasp fails, what is usually the real cause for humanoid robots such as the G1?

For vision-based grasping with humanoid robots such as the G1, the sources of failure in practice are listed one by one: camera depth or pose errors, hand-eye calibration drift or incorrect camera-to-hand transforms, latency from perception to control, and gripper issues (opening versus object size, friction, compliance, backlash). The poster is asking robotics grasping practitioners for feedback from real experience.