Offline Meta-Reinforcement Learning with Online Self-Supervision

Meta-reinforcement learning (RL) can be used to train policies that quickly adapt to new tasks with orders of magnitude less data than standard RL, but this fast adaptation often comes at the cost of greatly increasing the amount of reward supervision during meta-training time. Offline meta-RL removes the need to continuously provide reward supervision because rewards must only be provided once when the offline dataset is generated. In addition to the challenges of offline RL, a unique distribution shift is present in meta RL: agents learn exploration strategies that can gather the experience needed to learn a new task, and also learn adaptation strategies that work well when presented with the trajectories in the dataset, but the adaptation strategies are not adapted to the data distribution that the learned exploration strategies collect. Unlike the online setting, the adaptation and exploration strategies cannot effectively adapt to each other, resulting in poor performance. In this paper, we propose a hybrid offline meta-RL algorithm, which uses offline data with rewards to meta-train an adaptive policy, and then collects additional unsupervised online data, without any ground truth reward labels, to bridge this distribution shift problem. Our method uses the offline data to learn the distribution of reward functions, which is then sampled to self-supervise reward labels for the additional online data. By removing the need to provide reward labels for the online experience, our approach can be more practical to use in settings where reward supervision would otherwise be provided manually. We compare our method to prior work on offline meta-RL on simulated robot locomotion and manipulation tasks and find that using additional data and self-generated rewards significantly improves an agent's ability to generalize.