Inducing structure in reward learning by learning features

Reward learning enables robots to learn adaptable behaviors from human input. Traditional methods model the reward as a linear function of hand-crafted features, but that requires specifying all relevant features priori, which is impossible for real-world tasks. To get around this issue, recent deep Inverse Reinforcement Learning (IRL) rewards directly raw state challenging because robot has implicitly are important and how combine them, simultaneously. Instead, we propose divide-and-conquer approach: focus input specifically on separately, only then them into reward. We introduce novel type teaching an algorithm utilizes it complex space. The can using demonstrations, corrections, or other frameworks. demonstrate our method in settings where have be learned scratch, well some known. By first focusing feature(s), decreases sample complexity improves generalization over IRL baseline. show experiments with physical 7-DoF manipulator, user study conducted simulated environment.