Reviews: InfoGAIL: Interpretable Imitation Learning from Visual Demonstrations

Paper Summary: This paper focuses on using GANs for imitation learning using trajectories from an expert. The authors extend the GAIL (Generative Adversarial Imitation Learning) framework by including a term in the objective function to incorporate latent structure (similar to InfoGAN). The authors then proceed to show that using their framework, which they call InfoGAIL, they are able to learn interpretable latent structure when the expert policy has multiple modes and that in some setting this robustness allows them to outperform current methods. Paper Overview: The paper is generally well written. I appreciated that the authors first demon- started how the mechanism works on a toy 2D plane example before moving onto more complex driving simulation environment. This helped illustrate the core concepts of allowing the learned policy to be conditioned on a latent variable in a minimalistic setting before moving on to a more complex 3D driving simulation.