Autonomous Driving System Rule Learning Using Expert-Defined Causality

Abstract

An increasing number of road users are travelling freely in urban environments. Each of them has their own motion preferences but is expected to comply with the traffic laws. To cope with the motion discrepancies, autonomous vehicles require highly sophisticated reactive decision-making that can adapt their motion given the surrounding environment and the applicable traffic laws. Such decision-makers must be trustworthy, since each mistake can lead to a fatality, and performant, since they must estimate, at a high frequency, which behaviour to implement. This thesis describes how such functionality can be achieved with a practical rule engine learned from expert driving decisions and a precise notion of requirements.

We first demonstrate the feasibility of planning the motion of an autonomous vehicle by implementing a prototype that, given a curated training suite of driving examples, can create and maintain a two-layer rule-based theory. Assuming perfect perception, we then design a method that learns the rules based on a precise notion of requirements. An expert anticipates that the decision-maker can enter a state for which a requirement is unmet and therefore specifies with a set of template rules the cause of each anticipated violation. For each template rule, its antecedent entails a notion of causality, and its consequent specifies the behaviour to implement. The set of template rules are used as a labelling function. Namely, each time the decision-maker fails to satisfy a requirement, an associated template rule is used to address the misbehaviour. The rules of the rule-based theory are based on templates. The antecedent of such rules are automatically learned and may have been significantly altered to include new relevant constraints that are expected to cope better with the requirements.

Finally, considering that autonomous vehicles rely on sensor capabilities, we thereafter extend our method to compete in the Carla Leaderboard operational design domain. Using the same computer vision as the best performer for which there is code available, we demonstrate that our system can learn a policy that is explainable while performing better than our competitor on the set of provided requirements. This thesis has been divided into three phases, each of which strongly correlates with a paper submitted to a conference or journal for publication.

In the first phase, we assess the feasibility of the proposed rule-based architecture by implementing/deploying a rule engine prototype in a level-3 autonomous vehicle driving for 110 kilometres of field tests in an urban environment of the city of Waterloo. Namely, the prototype has an algorithm to create and iteratively refine a rule-based behaviour planner, using a two-layer rule-based theory. The first layer determines a set of feasible parametrized behaviours, given the perceived state of the environment. From these, a resolution function chooses the most conservative high-level maneuver. The second layer then reconciles the parameters into a single behaviour. Based on the set of traffic rules described in a driver handbook, an expert produces a set of training examples expressing the relevant change of behaviours. The algorithm presented in that paper performs hierarchical rule-based machine learning.

In the second phase, we formalize the construction of the training suite of driving examples that inevitably comes with the iterative development of the autonomous driving technology infrastructure of software and hardware. Namely, we explore how to extract knowledge from counterexamples encountered while driving in a city generated by the CARLA simulator. For that, we convert the requirements of the CARLA Autonomous Driving Leaderboard into a specification that is used to learn a rule-based policy. We assess the generalization of the learned rule-based policy by evaluating its performance on an unseen city generated by the same simulator. We then compare our performance with InterFuser, a state-of-the-art competing approach, and demonstrate that our method outperforms their method.

In the third phase, we use the computer vision and tracker of InterFuser and create our own path generator inspired by the route planner of TransFuser to demonstrate that our method can cope with sensor noise while achieving state-of-the-art performance. In this phase, we use the six official towns that form the CARLA Autonomous Driving Leaderboard as the training towns and attempt to generalize to two unseen towns. Although our initial goal was to become an entry on the CARLA Autonomous Driving Leaderboard, the evaluation infrastructure has become unavailable. Therefore, to be convincing that our approach achieves state-of-the-art performance, we create our own challenge by randomly generating novel routes both on the six official towns and two additional unseen towns that have been released by the same officials.

Although we demonstrate that our method outperforms a state-of-the-art end-to-end approach, we list in Limitations a number of issues that have not yet been addressed and constitute limitations to the results presented in this thesis. Thereafter, we speculate on how our method can be extended to mitigate some of these limitations.