Obstacle Avoidance for Autonomous Driving in CARLA Using Segmentation Deep Learning Models

Behavior Metrics is an open-sourced autonomous driving network comparison tool that allows the user to load and test their autonomous driving models in different scenarios and compare the performance metrics against other models. Currently, Behavior Metrics only supports the follow-the-line task, where the vehicle must drive along a circuit while maintaining proximity to the center of the lane, and provides multiple trained models for benchmarking. This project aims to expand the current stack by adding support for a route navigation task where the agent follows a sequence of high-level commands to reach a destination while avoiding obstacles in CARLA simulator, as well as providing an end-to-end learning solution for the task. The ultimate goal is a model that enables an ego vehicle to follow the route while avoiding collision with dynamic objects, such as pedestrians and other vehicles, and comprehensive evaluation metrics for the new task.