UC Berkeley

The ability to interact with other road participants is essential for autonomous vehicles in order to navigate under highly complex or critical driving scenarios. It is an extremely challenging task to enable autonomous vehicles interactively predict behaviors of others, and plan safe and high-quality motions for themselves. In fact, designing interaction-aware prediction and planning models is of great importance for autonomous vehicles to achieve full autonomy.

When designing a model, it is essential to consider our expectations for the model, which can be derived through two perspectives: 1) From the perspective of evaluation metrics (i.e. backward design); 2) From the perspective of task and user preferences (i.e. forward design). In this dissertation, a combination of forward and backward design process is proposed and it is argued that a model should be designed based on desired model properties inferred from metrics, tasks, and user preferences. A total of seven properties are covered in this dissertation: uncertainty, multi-modality, interpretability, flexibility, generalizability, reliability, and efficiency. The fundamental research question we addressed in this dissertation is: how to design interaction-aware prediction and planning models under different autonomous driving settings in order to achieve desired model properties.

This dissertation is divided into two parts. In Part I, we focus on the design of two-agent interaction models, which involves fundamental explorations of how certain model property can be achieved through algorithm design by taking the advantage of relatively low-complexity settings. Part II is concerned with the design of multi-agent interaction models, which includes comprehensive analysis and design for each desired model property under more complicated driving scenarios. In each part, we consider different prediction and planning related problems, and motivate our design by identifying desired model properties that correspond to each problem. By utilizing the proposed model design process under different problem settings, we demonstrate that our models are able to not only achieve desired properties, but also have great performances in terms of various evaluation criteria.