Abstract
Vehicle autonomy has the potential to bring many social benefits, such as improved
traffic safety and increased productivity. Modern autonomous vehicles are able to sense
their local environment, recognise relevant objects, and make driving decisions that obey
traffic rules. Nevertheless, many situations encountered during daily driving continue
to be challenging for autonomous vehicles, holding back the commercial deployment
of autonomous driving technology. In particular, motion planning in environments
that involve interactions with human drivers requires the design of algorithms that
can reason about the uncertain motion of other vehicles while relying on noisy and
incomplete sensor measurements. Given the stochasticity in human driving behaviour
and sensor limitations, effective handling of uncertainty is of paramount importance
for ensuring system safety and robustness.
This thesis makes several contributions towards enabling self-driving vehicles to
reason about the uncertain behaviour of other drivers and utilise this reasoning capability for planning. As our use case, we focus on the complex task of merging into
moving traffic where uncertainty can emanate from the behaviour of other drivers
and imperfect sensor measurements. We exploit the power of deep neural networks
in learning complex correlations from data for developing driver behaviour models.
We use these models for planning on two levels of abstraction: high-level, discrete
decisions that help the autonomous vehicle reach its destination safely and in a timely
manner, and low-level continuous actions that directly influence the vehicle’s dynamics.
For high-level planning, we propose an original driver model that combines domain
knowledge with modern deep learning, offering greater interpretability than black-box
models while producing predictions that maintain long-term accuracy. Further, we
use the model for planning via Monte Carlo tree search, where the long-term future
consequences of decisions are taken into consideration. For low-level planning, we
propose a sampling-based, model-predictive approach. Other contributions are made
towards learning strategies that improve the models’ predictive accuracy.