Context-Aware Autonomous Driving Using Meta-Reinforcement Learning

Reinforcement learning (RL) methods achieved major advances in multiple tasks surpassing human performance. However, most of RL strategies show a certain degree of weakness and may become computationally intractable when dealing with high-dimensional and non-stationary environments. In contrast, human are more comfortable with learning from little experience and adapting to unexpected perturbations. These differences are shaping the current research intending to guide agent policies and eschewing the above limits. In this paper, we build a meta-reinforcement learning (MRL) method embedding an adaptive neural network (NN) controller for efficient policy iteration in changing task conditions. Our main goal is to extend RL application to the challenging task of urban autonomous driving in CARLA simulator. The proposed approach yields higher performance and faster learning capabilities than conventionally pre-trained and randomly initialized RL algorithms.