Research map / Reinforcement learning / Policy gradient and actor-critic
Towards Safe Reinforcement Learning with a Safety Editor Policy
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a cooperative utility-maximizer and safety-editor policy instead of a single Lagrangian policy | Safe RL tasks · 14 safe RL tasks | SWU | Helped |
| Use two cooperating policies rather than a doubled-size single policy | SAC-actor2x-Lag · 14 safe RL tasks | SWU: 0.67 vs 1.10 | Helped |
| Use hinge loss on utility Q values instead of squared L2 action distance (ablation) | SEditor · SafeRacingObstacle | utility return | Helped |
| Use squared L2 action distance instead of utility-Q hinge loss (ablation) | SEditor · Safety Gym tasks | success rate | No clear effect |
| Replace additive editing with direct action overwrite (ablation) | SEditor · Safety Gym tasks | utility performance | Hurt |
| Use additive editing so the edited action stays close to the proposed action when possible (ablation) | Safety editor · Safety Gym tasks | utility performance | Helped |
| Train with a strict violation-rate target | SEditor · Safety Gym and Safe Racing | constraint violation rate: 5×10^-4 | Helped |
| Apply unconstrained SAC rather than SEditor | SEditor · CarButton1/2, CarPush2, CarGoal2 | constraint violation rate: up to 200 times less | Hurt |
| Train with the SEditor framework rather than SAC-actor2x-Lag | SEditor · Safety Gym, 10M steps | SWU: 0.89 -> 1.04 | Helped |
| Use SEditor rather than SAC on the unmodified PointGoal1 task | Safe RL · PointGoal1, 1×10^7 steps | utility score: 23 -> 27 | Helped |
Closest papers by meaning
- Safety Representations for Safer Policy Learning (ICLR 2025 · similarity 0.92)
- Latent Safety-Constrained Policy Approach for Safe Offline Reinforcement Learning (ICLR 2025 · similarity 0.92)
- Balance Reward and Safety Optimization for Safe Reinforcement Learning: A Perspective of Gradient Manipulation (AAAI 2024 · similarity 0.91)
- Safe Reinforcement Learning with Instantaneous Constraints: The Role of Aggressive Exploration (AAAI 2024 · similarity 0.91)
- Safeguarded Progress in Reinforcement Learning: Safe Bayesian Exploration for Control Policy Synthesis (AAAI 2024 · similarity 0.91)
- Enhancing Safety in Reinforcement Learning with Human Feedback via Rectified Policy Optimization (NeurIPS 2025 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.