Research map / Reinforcement learning / Policy gradient and actor-critic
PolicyFlow: Policy Optimization with Continuous Normalizing Flow in Reinforcement Learning
Shunpeng Yang, Ben Liu, Hua Chen
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 PPO Gaussian policy with PolicyFlow | PPO · IsaacLab G1 | terminal episodic reward: 25.4 -> 30.0 | Improved |
| replace PPO Gaussian policy with PolicyFlow | PPO · IsaacLab Navigation | terminal episodic reward: 3.5 -> 4.2 | Improved |
| replace PPO Gaussian policy with PolicyFlow | PPO · IsaacLab H1 | terminal episodic reward: 29.3 -> 27.3 | Hurt |
| replace PPO Gaussian policy with PolicyFlow | PPO · IsaacLab Lift-Cube | terminal episodic reward: 153.1 -> 154.6 | No clear effect |
| replace PPO Gaussian policy with PolicyFlow | PPO · IsaacLab Quadcopter | terminal episodic reward: 141.8 -> 141.0 | No clear effect |
| replace PPO Gaussian policy with PolicyFlow | PPO · IsaacLab Open-Drawer | terminal episodic reward: 99.8 -> 99.1 | No clear effect |
| replace PPO Gaussian policy with PolicyFlow | PPO · IsaacLab per-iteration training time | per-iteration time (ms): 43.0 -> 57.7 | Hurt |
| increase embedding dimensions to 512 for PolicyFlow | PolicyFlow · IsaacLab H1 training time | per-iteration time (ms): 63.4 -> 115.5 | Hurt |
| add Brownian regularizer to PolicyFlow (ablation) | PolicyFlow · MultiGoal / PointMaze exploration | goal coverage | Improved |
| add Brownian regularizer on top of Gaussian entropy regularization (w_g=0.001) (ablation) | PolicyFlow · MultiGoal | mode coverage | Improved |
| replace uniform noise injection (weight 0.05) with Brownian regularizer (ablation) | PolicyFlow · MultiGoal | mode coverage | Improved |
| use smaller clipping range (epsilon 0.1 instead of 0.2) (ablation) | PolicyFlow · IsaacLab ANYmal-D | episodic reward | Hurt |
Closest papers by meaning
- Normalizing Flows are Capable Models for Continuous Control (NeurIPS 2025 · similarity 0.89)
- Flow Matching Policy Optimization with Mirror Descent and Entropy Constraints (NeurIPS 2026 · similarity 0.88)
- Flow-DPPO: Divergence Proximal Policy Optimization for Flow Matching Models (NeurIPS 2026 · similarity 0.88)
- Random Policy Evaluation Uncovers Policies of Generative Flow Networks (ICML 2025 · similarity 0.87)
- Composite Flow Matching for Reinforcement Learning with Shifted-Dynamics Data (NeurIPS 2025 · similarity 0.87)
- ReinFlow: Fine-tuning Flow Matching Policy with Online Reinforcement Learning (NeurIPS 2025 · 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.
