Reinforcement Learning with Parameterized Actions via Online Abstractions

Researchers have developed a new reinforcement learning approach for parameterized action spaces, which involve both discrete action choices and continuous parameters. Existing methods struggle: planning requires hand-crafted models, and standard RL algorithms handle only one type. The proposed method enables agents to autonomously learn state and action abstractions online, progressively refining them during learning. This extends RL to long-horizon, sparse-reward settings. The work is published on arXiv (2512.20831).