Interpretable Planning and Prediction for Autonomous Vehicles

Accurately predicting the intentions and actions of other road users and then using this information during motion planning is an important task in the field of autonomous vehicles. However, current state of the art methods fail to match human performance in many ways, making mistakes for situations that would be trivial for humans to handle. Many current methods also use black box techniques such as deep neural networks, which is not ideal for a safety critical task such as driving. This leaves an open question : Which planning and prediction methods can approach or surpass human driving performance while also maintaining interpretability?


During this project I will develop new prediction and planning methods, some of which may build upon previous methods. Initially I plan to using decision trees to perform goal recognition in order to make predictions about other vehicles, and then integrate this with a Monte Carlo Tree Search (MCTS) planner. This method would identify the goal of other vehicles, such as taking a certain exit at a junction, and then use the inferred goal to make predictions about the vehicle's future behaviour. I will compare methods of designing the decision trees by hand, and training them from data. This data could be obtained either through simulation or recordings of real vehicles, for which there are several open datasets available.


Later in the project I could investigate other prediction and planning methods. One recent area of research has been "Programmatically Interpretable Reinforcement Learning" (PIRL) which is designed to generate interpretable and verifiable agent policies. I could investigate whether such methods can be adapted to generate policies for realistic autonomous driving. Methods developed during the course of this project could also have the potential to generalise to robotics domains other than autonomous driving.