Abstract

Many of the successes in deep learning build upon rich supervision. Reinforcement learning (RL) is no exception to this: algorithms for locomotion, manipulation, and game playing often rely on carefully crafted reward functions that guide the agent. But defining dense rewards becomes impractical for complex tasks. Moreover, attempts to do so frequently result in agents exploiting human error in the specification. To scale RL to the next level of difficulty, agents will have to learn autonomously in the absence of rewards. We define task-agnostic reinforcement learning (TARL) as learning in an environment without rewards to later quickly solve down-steam tasks. Active research questions in TARL include designing objectives for intrinsic motivation and exploration, learning unsupervised task or goal spaces, global exploration, learning world models, and unsupervised skill discovery. The main goal of this workshop is to bring together researchers in RL and investigate novel directions to learning task-agnostic representations with the objective of advancing the field towards more scalable and effective solutions in RL.