Reinforcement Learning without Ground-Truth State

摘要

To perform robot manipulation tasks, a low-dimensional state of the environment typically needs to be estimated. However, designing a state estimator can sometimes be difficult, especially in environments with deformable objects. An alternative is to learn an end-to-end policy that maps directly from high-dimensional sensor inputs to actions. However, if this policy is trained with reinforcement learning, then without a state estimator, it is hard to specify a reward function based on high-dimensional observations. To meet this challenge, we propose a simple indicator reward function for goal-conditioned reinforcement learning: we only give a positive reward when the robot's observation exactly matches a target goal observation. We show that by relabeling the original goal with the achieved goal to obtain positive rewards (Andrychowicz et al., 2017), we can learn with the indicator reward function even in continuous state spaces. We propose two methods to further speed up convergence with indicator rewards: reward balancing and reward filtering. We show comparable performance between our method and an oracle which uses the ground-truth state for computing rewards. We show that our method can perform complex tasks in continuous state spaces such as rope manipulation from RGB-D images, without knowledge of the ground-truth state.