safe-reinforcement-learning

本文提出 RBGA（Residual-Budget Guide Arbitration）算法，通过显式跟踪回合内剩余安全预算，在离线引导策略与在线策略之间进行逐步仲裁，实现离线到在线的安全强化学习，并在灵巧操作任务中验证了其在线性能提升与安全保障。

本文提出了一种物理信息图表示与约束专用评论家相结合的安全深度强化学习框架（LDCPPO），用于配电网中社区电池储能的协调控制，以缓解电压和线路越限问题并降低运行成本，实验显示关键电压违规程度最高降低80.5%。

FailBank, a four-stage self-evolving framework, converts runtime safety-shield feedback from a CBF-based teacher into persistent policy improvements via guarded LoRA updates, improving task success by up to 25.4 percentage points over runtime shielding on the VLA-Arena benchmark.

This paper introduces certified safety curation for safe offline reinforcement learning, using a filter-then-clone pipeline with distribution-free guarantees to select safe trajectories without needing cost functions on every transition.

This paper introduces adaptive probabilistic shielding for safe reinforcement learning, where the shield is computed from an online learned MDP model, adapting as the model becomes more accurate.

SteinGate introduces a distributional safety certificate using Kernelized Stein Discrepancy to detect rare catastrophic tail events in safe reinforcement learning, dynamically adapting policy updates to reduce constraint violations while maintaining competitive returns.

This work-in-progress paper proposes embedding a differentiable physics model into the PPO actor loss function to penalize anticipated safety violations in reinforcement learning, evaluated on a simulated 1-DoF helicopter system. The physics-informed soft regularizations reduce constraint violations while maintaining reliable target tracking.

This paper proposes Constraint-Sensitive Policy Optimization (CSPO), a first-order primal-dual method for safe reinforcement learning that incorporates local constraint sensitivity to improve safety recovery and reduce oscillations near safety boundaries, achieving higher constrained returns on navigation and locomotion benchmarks.

A method for contract-based compositional shielding that ensures global safety in multi-agent reinforcement learning without centralized runtime control, using local LTL obligations and a multi-armed bandit to optimize team reward.

Introduces a novel shielding framework for robust Markov decision processes (RMDPs) that formally guarantees safety under uncertain transition dynamics, proving soundness and optimality. The approach combines with PAC guarantees for learned models, enabling safe reinforcement learning in unknown environments.