IEEE AESS: Exploring Reinforcement Learning Techniques for Autonomous Vehicle Control Using the CARLA Simulator

Reinforcement learning has gained traction in research as a machine learning method for the control of autonomous vehicles. Its paradigm of learning by traversing through a series of states has shown to work well with the dynamic nature of driving. However, for safety reasons, much of the research effort is performed in simulators. This work sought to utilize a driving simulator to train an autonomous vehicle using various reinforcement learning methods, with the goal of the vehicle to reach a destination while staying within the road and avoiding collisions with other obstacles.
Three methods were explored. The first utilized a Deep Q-Network (DQN) based algorithm with the state being defined by direct values and reward shaping. The second method extended upon the first, utilizing a Proximal Policy Optimization (PPO) based algorithm. It included additions to the reward function based on lessons learned and experimentation with a CNN based state representation. It was shown to learn to solve a driving scenario with less needed steps, as compared to the first method. The final method utilized behavioral cloning with images as its state representation. It was shown to be a minor improvement over its baseline.
Talk is limited to US citizens. Registration is required for non-SwRI employees.
Cookies and drinks will be provided.
Speaker(s): Joseph Clemmons
Bldg: Building 51, 6220 Culebra Rd, San Antonio, Texas, United States, 78238