Prioritized offline Goal-swapping Experience Replay
This addresses data scarcity in offline goal-conditioned RL for robotics and manipulation tasks, but it is incremental as it builds on existing goal-swapping methods.
The paper tackled the challenge of generating valid additional data in goal-conditioned offline reinforcement learning by proposing prioritized goal-swapping experience replay (PGSER), which uses a pre-trained Q function to prioritize goal-swapped transitions that allow reaching the goal, resulting in significant improvements over baselines across benchmark tasks, including dexterous in-hand manipulation.
In goal-conditioned offline reinforcement learning, an agent learns from previously collected data to go to an arbitrary goal. Since the offline data only contains a finite number of trajectories, a main challenge is how to generate more data. Goal-swapping generates additional data by switching trajectory goals but while doing so produces a large number of invalid trajectories. To address this issue, we propose prioritized goal-swapping experience replay (PGSER). PGSER uses a pre-trained Q function to assign higher priority weights to goal swapped transitions that allow reaching the goal. In experiments, PGSER significantly improves over baselines in a wide range of benchmark tasks, including challenging previously unsuccessful dexterous in-hand manipulation tasks.