Duckietown environment

The dataset used in this paper is a collection of state-action pairs generated by a pre-trained RL agent, used to train a self-supervised interpretable network (SSINet) to produce attention masks for explaining the agent's decisions.