Current self-driving technology may exacerbate traffic congestion

While the stop-and-go traffic has been a long-standing concern for traffic flow researchers, it has not received enough attention from the self-driving industry, partially due to the gap between the two separate domains.

This small post attempts to identify a research gap between the AV industry and traffic congestion.

Comma.ai’s openpilot (OP) is one of the cutting-edge self-driving models available to everyone. With such special interest in congestion, or stop-and-go traffic, we test the model performance on how well it can predict the future speed.

Speed prediction when the vehicle is fully stopped in congestion

One can see the OP model outputs a wrong prediction on the future speed when the vehicle is actually fully stopped in congestion.

Speed prediction in stop-and-go traffic

We suggest learning an unbiased self-driving model from more evenly-distributed data. We prepared a dataset with the same amount of driving data for different speed groups ranging from [0,5], [5,10]…[20,25] (m/s).

Learning an unbiased self-driving model