Research map / Reinforcement learning / Policy gradient and actor-critic

Feasible Policy Optimization for Safe Reinforcement Learning

NeurIPS 2026 · Reinforcement learning › Policy gradient and actor-critic · approach: Safety layers and shielding

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's abstract only. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a safe RL policy optimization approach that relaxes per-update constraint enforcement, progressively expanding the feasible region while improving value.

Problem

Safe reinforcement learning requires policies to strictly satisfy safety constraints, but enforcing constraints at every policy update is viewed as overly conservative and hard to achieve with policy gradient methods.

Approach

Argues that each policy update only needs to progressively expand the feasible region while improving the value function, rather than satisfying constraints at every update, yielding a less conservative policy gradient method for safe RL.

Surprising findings

• Enforcing safety constraints at every policy update is unnecessarily conservative.

Closest papers by meaning

• Enhancing Efficiency of Safe Reinforcement Learning via Sample Manipulation (NeurIPS 2024 · similarity 0.91)
• Constrained Reinforcement Learning Under Model Mismatch (ICML 2024 · similarity 0.90)
• Safe Reinforcement Learning with Instantaneous Constraints: The Role of Aggressive Exploration (AAAI 2024 · similarity 0.90)
• P2BPO: Permeable Penalty Barrier-Based Policy Optimization for Safe RL (AAAI 2024 · similarity 0.90)
• Embedding Safety into RL: A New Take on Trust Region Methods (ICML 2025 · similarity 0.90)
• Off-Policy Primal-Dual Safe Reinforcement Learning (ICLR 2024 · similarity 0.90)