[Paper] Is this the real driver, or is it just a robot?

After developing DriViDOC (https://lnkd.in/eEHUyGa6), a controller that combines differentiable MPC with end-to-end imitation learning from images, I wanted to go beyond objective metrics and ask: do people actually perceive it as more human-like than a standard MPC?

The answer was surprisingly clear. Even though both controllers completed the same lane-keeping task correctly, participants could barely distinguish DriViDOC from a human driver. The baseline MPC, despite being an advanced control tool, was spotted almost every time. And the reasons were subtle: how the car positioned itself within the lane, small variations in speed profile. Not dramatic failures, just the quiet and nuanced behavioral patterns that makes driving feel human.

We also introduce an evaluation framework, based on a Turing-type test on a moving-base simulator, that I hope will be useful for future design and validation of human-like AV systems.

A huge thank you to my co-authors and PhD supervisors Jan Swevers, Tinne Tuytelaars, and Son Tong, and to all 40 participants who sat on a Stewart platform and patiently let themselves be driven by robots 🤖

Full paper free access through the link below, enjoy the read!

Article Link