Research map / Reinforcement learning / Offline RL
Geometric Control of Out-of-Distribution Shift in Safe Offline RL
Zhiqi Zhuang, Di Wu, Benoit Boulet
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| replace HJ reachability feasibility with cost-threshold feasibility removed (add HJ) (ablation) | GSCO · Bullet-Safety-Gym BallRun | normalized cost: 6.56 -> 0.00 | Improved |
| add HJ reachability feasibility (ablation) | GSCO · Bullet-Safety-Gym DroneCircle | normalized cost: 0.67 -> 0.10 | Improved |
| add HJ reachability feasibility (ablation) | GSCO · Bullet-Safety-Gym CarRun | normalized reward: 0.95 -> 0.84 | Hurt |
| use conditional flow prior instead of Gaussian prior (ablation) | GSCO · Safety-Gymnasium (6 tasks) | normalized reward and cost: e.g. CarButton1 cost 1.54 -> 0.15 | Mixed |
| add refinement (any order) over sampling from flow prior (ablation) | GSCO · Bullet-Safety-Gym (4 tasks) | normalized return | Improved |
| order safety refiner before reward refiner (H->R->SH vs R->H->SH) (ablation) | GSCO · Bullet-Safety-Gym (4 tasks) | normalized cost: lower cost, slightly lower return | Mixed |
| use random refiner order (ablation) | GSCO · Bullet-Safety-Gym (4 tasks) | normalized return | Hurt |
| increase refinement steps T from 3 to larger (ablation) | GSCO · CarCircle | stability | Hurt |
| raise expectile tau_h of feasibility critic (ablation) | HJ feasibility classifier · CarRun | recall: 0.32 -> 0.85 (precision 0.76 -> 0.21) | Improved |
| add geometric regularization L_reg (ablation) | GSCO · BallCircle, CarPush1, MediumDense | Gaussian-fit KL and sliced W2 | Improved |
| use GSCO instead of FISOR and LSPC | GSCO · MediumDense | OOD rate: 0.113 (FISOR), 0.086 (LSPC) -> 0.033 | Improved |
| use GSCO instead of FISOR | GSCO · Safety-Gymnasium avg | normalized cost: 0.40 -> 0.31 | Improved |
Closest papers by meaning
- Relaxed Stationary Distribution Correction Estimation for Improved Offline Policy Optimization (AAAI 2024 · similarity 0.88)
- Latent Safety-Constrained Policy Approach for Safe Offline Reinforcement Learning (ICLR 2025 · similarity 0.88)
- Selective Uncertainty Propagation in Offline RL (AAAI 2025 · similarity 0.88)
- Improving Generalization in Offline Reinforcement Learning via Latent Distribution Representation Learning (AAAI 2025 · similarity 0.87)
- Robust Offline Reinforcement Learning with Linearly Structured $f$-Divergence Regularization (ICML 2025 · similarity 0.87)
- Iteratively Refined Behavior Regularization for Offline Reinforcement Learning (NeurIPS 2024 · similarity 0.87)
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.
