Autonomous Driving in Old University Towns
AbstractAutonomous driving simulators are paramount to the development of effective policies to drive autonomous cars. For example, the popular open-source CARLA simulator, implemented using Unreal Engine, has driven the community forward by providing open digital assets simulating american towns of various characteristics. However, we would expect autonomous car agents trained in an american town simulator to fail when deployed in new scenarios, for example when deployed in old university towns. This is because many visual scenes are very different from what the agent saw at training time, thus the agent would not know how to react. In this project we will try to validate this hypothesis by extending the current assets implemented in the simulator to include Old University Towns, and assess the robustness of imitative models trained on american towns, when deployed in Old University Towns. We will use Google Maps to create a new map for the simulator based on Oxford city centre, and use Google Street View to create new assets for the simulator as well.
Requirements * strong python experience * experience with deep learning, including computer vision models * preferred: experience with unreal engine and asset creation