Abstract
One of the biggest challenges in the development of
learning-driven automated driving technologies remains the
handling of uncommon, rare events that may have not been
encountered in training. Especially when training a model with
real driving data, unusual situations, such as emergency
brakings, may be underrepresented, resulting in a model that
lacks robustness in rare events. This study focuses on
car-following based on reinforcement learning and demonstrates
that existing approaches, trained with real driving data, fail to
handle safety--critical situations. Since collecting data
representing all kinds of possible car-following events,
including safety--critical situations, is challenging, we propose
a training environment that harnesses stochastic processes to
generate diverse and challenging scenarios. Our experiments show
that training with real data can lead to models that collide in
safety--critical situations, whereas the proposed model exhibits
excellent performance and remains accident-free, comfortable, and
string-stable even in extreme scenarios, such as full-braking by
the leading vehicle. Its robustness is demonstrated by simulating
car-following scenarios for various reward function
parametrizations and a diverse range of artificial and real
leader data that were not included in training and were
qualitatively different from the learning data. We further show
that conventional reward designs can encourage aggressive
behavior when approaching other vehicles. Additionally, we
compared the proposed model with classical car-following models
and found it to achieve equal or superior results.
Links and resources
Tags
