Research map / Reinforcement learning / Policy gradient and actor-critic
Safety layers and shielding
Add safety editors, shields, or constraint-conditioned policies to keep agents within safe behavior.
85 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Policy gradient and actor-critic, keyword cluster “safe · constrained · constraints”
State History Tried Baselines Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
31 papers in 2023–24 and 29 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 9.0% to 6.3%.
Open questions the data points to
- Papers report PPO-Lag on PointGoal1 (reward) anywhere from 13.4 to 15: comparisons against it depend on whose number you take.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2017 · Safe Model-based Reinforcement Learning with Stability Guarantees (NeurIPS 2017 · 319 citations)
- 2018 · Safe Reinforcement Learning via Shielding (AAAI 2018 · 687 citations)
- 2019 · End-to-End Safe Reinforcement Learning through Barrier Functions for Safety-Critical Continuous Control Tasks (AAAI 2019 · 575 citations)
- 2020 · Responsive Safety in Reinforcement Learning by PID Lagrangian Methods (ICML 2020 · 43 citations)
- 2021 · WCSAC: Worst-Case Soft Actor Critic for Safety-Constrained Reinforcement Learning (AAAI 2021 · 100 citations)
- 2022 · Exploring Safer Behaviors for Deep Reinforcement Learning (AAAI 2022 · 27 citations)
- 2023 · Evaluating Model-Free Reinforcement Learning toward Safety-Critical Tasks (AAAI 2023 · 23 citations)
- 2024 · Balance Reward and Safety Optimization for Safe Reinforcement Learning: A Perspective of Gradient Manipulation (AAAI 2024 · 12 citations)
- 2025 · Tilted Quantile Gradient Updates for Quantile-Constrained Reinforcement Learning (AAAI 2025 · 0 citations)
- 2026 · CSPO: Constraint-Sensitive Policy Optimization for Safe Reinforcement Learning (ICML 2026 · 0 citations)
Earliest papers here
- Safe Model-based Reinforcement Learning with Stability Guarantees (NeurIPS 2017 · 319 citations)
- Safe Reinforcement Learning via Shielding (AAAI 2018 · 687 citations)
- Diverse Exploration for Fast and Safe Policy Improvement (AAAI 2018 · 10 citations)
What has been tried
65 papers here report 520 measured changes (a component added, removed or swapped, compared with a baseline): helped 64%, no clear effect 7%, hurt 23%, mixed 6%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Use PCRPO-2SR slack settings instead of CRPO | SafetyWalker · reward and safety balance | Helped | Balance Reward and Safety Optimization for Safe Reinforcement Learning: A Perspective of Gradient Manipulation (AAAI 2024 · 12 citations) | |
| Use only good demonstrations, removing bad demonstrations | SafetyGym (two environments) · expected reward and safety | Hurt | Imitate the Good and Avoid the Bad: An Incremental Approach to Safe Reinforcement Learning (AAAI 2024 · 2 citations) | |
| Vary the permeability parameter β from 1 to 10, 15, or 100 | Humanoid Run · reward | Mixed | P2BPO: Permeable Penalty Barrier-Based Policy Optimization for Safe RL (AAAI 2024 · 2 citations) | |
| Add incremental verification | Lane Following, 20-step training · Runtime (seconds; 20 epochs) | 185.5 -> 24.8 | Helped | Verified Safe Reinforcement Learning for Neural Network Dynamic Models (NeurIPS 2024 · 2 citations) |
| add tilted update to QCPO | SimpleEnv 95% and DynamicEnv 95% · Return | 2.1±0.2 -> 4.7±0.2; 3.5±0.1 -> 8.0±0.3 | Helped | Tilted Quantile Gradient Updates for Quantile-Constrained Reinforcement Learning (AAAI 2025 · 0 citations) |
| add the modified trust region to the method | Safety Gymnasium tasks · safety violations | Helped | Embedding Safety into RL: A New Take on Trust Region Methods (ICML 2025 · 0 citations) | |
| Set uncertainty level K=1 instead of an intermediate level | Safe-control-gym tasks, observation/action/dynamics uncertainty · episodic risk | Hurt | Fuz-RL: A Fuzzy-Guided Robust Framework for Safe Reinforcement Learning under Uncertainty (NeurIPS 2025 · 0 citations) | |
| Add model-uncertainty pessimism to exploration | Pendulum and Cartpole exploration · constraint violations during learning | Helped | ActSafe: Active Exploration with Safety Constraints for Reinforcement Learning (ICLR 2025 · 0 citations) | |
| Increase the latent safety restriction parameter | restriction-parameter evaluation · reward return | Helped | Latent Safety-Constrained Policy Approach for Safe Offline Reinforcement Learning (ICLR 2025 · 0 citations) | |
| Use expected failure likelihood for the current policy instead of a state-centric steps-to-cost distribution | Experiments shown in Alternative Safety Representations · Task performance | No clear effect | Safety Representations for Safer Policy Learning (ICLR 2025 · 0 citations) | |
| Use OSRP instead of BSRP-Lag | SafetyPointButton1 · reward | Helped | SafeDreamer: Safe Reinforcement Learning with World Models (ICLR 2024 · 0 citations) | |
| Add cost conservatism by setting k above 0 | PointButton2 and CarButton1 · constraint satisfaction | Helped | Off-Policy Primal-Dual Safe Reinforcement Learning (ICLR 2024 · 0 citations) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| PPO-Lag | PointGoal1 | reward | 13.4 – 15 | 2 papers, e.g. ICML 2024 and AAAI 2024 |
What reviewers will expect
Datasets and baselines named in the 18 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Safety-Gym · 2 of 18 papers
Baselines
- PPO-Lag · 4 of 18 papers
- CPO · 4 of 18 papers
- PCPO · 2 of 18 papers
- CUP · 2 of 18 papers
- DQN · 2 of 18 papers
- CPQ · 2 of 18 papers
- CDT · 2 of 18 papers
- PPO-Lagrangian · 2 of 18 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 60 papers; from OpenAlex and the papers' affiliations).
- Carnegie Mellon University · 6 papers
- Peking University · 5 papers
- Zhejiang University · 4 papers
- Tsinghua University · 4 papers
- University of California, Berkeley · 3 papers
- Virginia Tech · 3 papers
- Microsoft · 2 papers
- Washington University · 2 papers
- Massachusetts Institute of Technology · 2 papers
- Google · 2 papers
Use it
Download BibTeX of 39 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Constrained policy optimization (48 papers)
- Safe policy improvement (27 papers)
Elsewhere in the topic
- Model-based RL with guarantees (245 papers)
- Regularized scalable deep value learning (216 papers)
- Temporal-difference and discounting theory (143 papers)
- Tree search and policy optimization (109 papers)
- Variance-reduced policy gradients (79 papers)
- Actor-critic theory (76 papers)
- Off-policy actor-critic with expressive policies (71 papers)
- Trajectory reuse in policy gradients (68 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).