Researchers have developed a method for safe reinforcement learning called adaptive probabilistic shielding. In traditional reinforcement learning, an observer called the shield constrains the agent's actions to those that are safe based on pre-known transition probabilities of the underlying Markov decision process (MDP). However, this approach is not applicable when the MDP model is unknown. The new method integrates probabilistic shielding with online model learning, where the transition probabilities are estimated as the agent explores the environment. This allows for a shield that adapts to the changing estimates and improves in tandem with the agent's learning. --- Why it matters: This matters because it enables safe reinforcement learning in scenarios where the underlying MDP is unknown or partially known, which is common in real-world applications. The method can improve the safety of autonomous systems and robots by adapting to changing environments. Source: https://arxiv.org/abs/2608.19836