Research map / Reinforcement learning / Visual and representation RL
Network Randomization: A Simple Technique for Generalization in Deep Reinforcement Learning
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 |
|---|---|---|---|
| Randomize input observations with a random network | PPO · CoinRun small-scale, unseen | success rate (%): 34.6 -> 76.7 | Helped |
| Add feature-matching loss to randomized-input training (ablation) | PPO + network randomization · CoinRun small-scale, unseen | success rate (%): 76.7 -> 78.1 | Helped |
| Use randomized inputs without feature-matching loss (ablation) | PPO + network randomization + feature matching · CoinRun small-scale, unseen | success rate (%): 78.1 -> 76.7 | Hurt |
| Replace network randomization with cutout augmentation | PPO · CoinRun small-scale, unseen | success rate (%): 78.1 -> 41.9 | Hurt |
| Replace network randomization with color jitter | PPO · CoinRun small-scale, unseen | success rate (%): 78.1 -> 43.1 | Hurt |
| Add network randomization and feature matching | PPO · CoinRun small-scale, unseen | 2-way cycle-consistency (%): 18.9 -> 67.8 | Helped |
| Add network randomization and feature matching | PPO · CoinRun small-scale, unseen | 3-way cycle-consistency (%): 4.4 -> 43.3 | Helped |
| Apply network randomization instead of cutout | PPO · CoinRun large-scale, unseen themes | success rate (%): 39.8% to 58.7% | Helped |
| Randomize network inputs during training | PPO · DeepMind Lab, unseen environments | total rewards: 55.4 -> 358.2 | Helped |
| Randomize network inputs during training | PPO · Surreal robotics, unseen environments | total rewards: 59.2 -> 356.8 | Helped |
| Increase the number of Monte Carlo samples at inference (ablation) | PPO + network randomization · CoinRun small-scale, unseen | success rate | Helped |
| Randomize the first network layer rather than higher layers (ablation) | PPO + network randomization · CoinRun large-scale, unseen | performance | Helped |
Closest papers by meaning
- Quantifying Generalization in Reinforcement Learning (ICML 2019 · similarity 0.90)
- Random Latent Exploration for Deep Reinforcement Learning (ICML 2024 · similarity 0.87)
- Flow-based Domain Randomization for Learning and Sequencing Robotic Skills (ICML 2025 · similarity 0.87)
- Learning Randomized Algorithms with Transformers (ICLR 2025 · similarity 0.86)
- Regularized Parameter Uncertainty for Improving Generalization in Reinforcement Learning (CVPR 2024 · similarity 0.86)
- Domain Randomization via Entropy Maximization (ICLR 2024 · similarity 0.86)
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.