Core Viewpoint - The article emphasizes the shift in focus for autonomous driving from "solving problems" to "defining problems," highlighting the importance of evaluation over training in the latter stage of development [1]. Group 1: RL Prior - RL prior refers to the existing knowledge or assumptions about a task before new data is observed, which is crucial for guiding the learning process of an agent [3]. - The core function of RL prior is to provide an "initial starting point" for the agent, preventing it from starting as a blank slate [3]. - RL prior can be obtained through various methods, including human expert experience, domain knowledge, transfer learning, offline data, and task decomposition [5][6]. Group 2: Benefits of RL Prior - The benefits of using RL prior include reducing the learning cost, which manifests in three main aspects: constraining the exploration space, optimizing model initialization, and shaping the reward function [7][12]. - By incorporating prior knowledge, the agent can avoid random actions and focus on reasonable exploration, leading to faster convergence and reduced sample complexity [13][14]. Group 3: Application of RL Prior - In robotics, RL prior can significantly lower training difficulty by integrating physical constraints and expert knowledge, such as optimizing trajectories for robotic arms and navigation for mobile robots [10][11]. - In gaming, RL prior can enhance performance by utilizing game rules and human player experiences, enabling agents to surpass human levels quickly [16][17]. - In recommendation systems, RL prior helps address cold start and reward delay issues by leveraging user behavior history and item attributes [19][20].