Human-Machine Co-Driving's 'Reviewer' Dilemma: End-to-End Needs Better Experiment Design

Every version iteration of an autonomous driving system is actually like revising a paper—users are the reviewers, crashes are rejection letters, and NIO's Onvo pushing Coconut 3.1.0 this time feels more like submitting a revised version with "supplementary experiments." From the perspective of a computer vision researcher, the biggest highlight of this upgrade isn't how many features were added, but the engineering implementation of the concept of "human-machine co-driving mode." It touches on a core issue that has been debated in academia for years: when end-to-end models start taking over the perception-decision pipeline, what role should human drivers play?