PDRL: Towards Deeper States and Further Behaviors in Unsupervised Skill Discovery by Progressive Diversity

摘要

We present progressive diversity reinforcement learning (PDRL), an unsupervised reinforcement learning (URL) method for discovering diverse skills. PDRL encourages learning behaviors that span multiple steps, particularly by introducing “deeper states”—states that require a longer sequence of actions to reach without repetition. To address the challenges of weak skill diversity and weak exploration in partially observable environments, PDRL employs two indications for skill learning to foster exploration and skill diversity, emphasizing each observation and subtrajectory's accuracy compared to its predecessor. Skill latent variables are represented by mappings from states or trajectories, helping to distinguish and recover learned skills. This dual representation promotes exploration and skill diversity without additional modeling or prior knowledge. PDRL also integrates intrinsic rewards through a combination of observations and subtrajectories, effectively preventing skill duplication. Experiments across multiple benchmarks show that PDRL discovers a broader range of skills compared to existing methods. Additionally, pretraining with PDRL accelerates fine-tuning in goal-conditioned reinforcement learning (GCRL) tasks, as demonstrated in Fetch robotic manipulation tasks.