To operate autonomously, a robotic vehicle must perform actions to successfully complete a specified task. To reliably plan for scenarios with significant uncertainty (due to factors like unknown and dynamic environments, noisy sensing, and the actions of external agents), the problem is often modelled as a sequential decision-making problem. Reinforcement learning (RL) is a popular technique from the field of computer science that can be used to solve such decision-making problems. RL has seen growing popularity in recent years; successful applications include agents capable of beating grandmaster-level human players at the games of Chess or Go.
In most of these cases, however, model-free RL techniques were used. This trial-and-error-based approach refers to algorithms where the agent learns the optimal behaviour directly from experience. Though simulation-based setups are forgiving in this regard, such naive exploration could lead to disastrous consequences and irreparable damage in practical robotic systems. A more viable approach is therefore to use model-based RL. This would enable the agent to use its experience to explicitly construct (or improve) a model of its environment. It could then use the latest model to plan and predict the outcomes of actions – something that will lead to much safer operation in practice. The aim of this project is to develop such a model-based RL framework that can be used to solve a set of autonomous driving tasks. This will first be done in simulation before applying the solution to a physical vehicle.