Beyond RMSE: Do Machine-Learned Models of Road User Interaction Produce Human-Like Behavior?

Autonomous vehicles use a variety of sensors and machine-learned models to predict the behavior surrounding road users. Most in literature focus on quantitative error metrics like root mean square (RMSE) learn report their models' capabilities. This tends ignore more important behavioral aspect models, raising question whether these really human-like behavior. Thus, we propose analyze output much would human data conventional research. We introduce demonstrate presence three different phenomena naturalistic highway driving dataset: 1) The kinematics-dependence who passes merging point first 2) Lane change by an on-highway vehicle accommodate on-ramp 3) changes avoid lead conflicts. Then, using same metrics. Even though RMSE value differed, all captured kinematic-dependent but struggled at varying degrees capture nuanced courtesy lane Additionally, collision aversion analysis during showed that physical driving: leaving adequate gap between vehicles. our highlighted inadequacy simple need take broader perspective when analyzing predictions.