Research map / Reinforcement learning / Policy gradient and actor-critic
Image Augmentation Is All You Need: Regularizing Deep Reinforcement Learning from Pixels
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 |
|---|---|---|---|
| Add random image shifts to SAC (ablation) | SAC trained from pixels · DeepMind Control Suite | episode return | Helped |
| Add target-Q averaging over transformed observations (K>1) (ablation) | DrQ / SAC · DeepMind Control Suite | episode return | Helped |
| Add Q-function averaging over transformed observations (M>1) (ablation) | DrQ / SAC · DeepMind Control Suite | episode return | Helped |
| Increase K and M augmentation counts (ablation) | DrQ / SAC · DeepMind Control Suite, especially Cheetah Run | episode return | Helped |
| Use random shifts rather than the other tested image augmentations (ablation) | SAC image augmentation · DeepMind Control Suite | episode return | Helped |
| Perform four updates per environment step instead of one (ablation) | DrQ · DeepMind Control Suite, 100k environment steps | sample efficiency | Helped |
| Perform four updates per environment step instead of one (ablation) | SAC without augmentation · DeepMind Control Suite | overfitting | Hurt |
| Add DrQ image augmentation (random shifts and intensity) | Efficient DQN · Atari 100k, 26 games | median human-normalized episode return: 0.058 -> 0.268 | Helped |
| Apply DrQ to Efficient DQN | Atari 100k · Atari 100k, 26 games | median human-normalized episode return: 0.058 -> 0.268 | Helped |
| Use DrQ [K=2,M=2] rather than the strongest listed pixel-based baseline | SAC · PlaNet benchmark, 100k steps | episode return | Mixed |
| Use DrQ [K=2,M=2] rather than the strongest listed pixel-based baseline | SAC · PlaNet benchmark, 500k steps | episode return | Helped |
Closest papers by meaning
- Revisiting Data Augmentation in Deep Reinforcement Learning (ICLR 2024 · similarity 0.90)
- Stabilizing Deep Q-Learning with ConvNets and Vision Transformers under Data Augmentation (NeurIPS 2021 · similarity 0.90)
- Fourier Guided Adaptive Adversarial Augmentation for Generalization in Visual Reinforcement Learning (AAAI 2025 · similarity 0.85)
- VIGOR: Zero-Shot Visual Generalization via Latent-Space Consistency in Model-Based Reinforcement Learning (NeurIPS 2026 · similarity 0.85)
- Zero Shot Generalization of Vision-Based RL Without Data Augmentation (ICML 2025 · similarity 0.85)
- Efficient Reinforcement Learning Through Adaptively Pretrained Visual Encoder (AAAI 2025 · similarity 0.84)
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.