超越硬约束：预算条件下的安全离线强化学习可达性

Sequential decision-making using Markov Decision Processes (MDPs) is foundational in numerous real-world applications. Both model-based and model-free methods have demonstrated strong performance in these contexts. However, real-world tasks often necessitate a balance between reward maximization and safety constraints, which can lead to conflicting objectives and unstable min/max adversarial optimization. Safety reachability analysis emerges as a promising alternative, precomputing a forward-invariant safe state and action set that ensures an agent starting within this set remains safe indefinitely. Nevertheless, most reachability-based methods focus solely on hard safety constraints, with limited work extending reachability to cumulative cost constraints. To address this gap, a safety-conditioned reachability framework is defined, aiming to extend existing reachability analysis methods to encompass safety under budget conditions.