Research map / Reinforcement learning / Visual and representation RL
Stabilizing Deep Q-Learning with ConvNets and Vision Transformers under Data Augmentation
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 |
|---|---|---|---|
| Use SVEA rather than DrQ with ViT | SAC with ViT encoder · Walker, walk color_hard | episode return: 94 -> 877 | Helped |
| Use ViT rather than CNN with SVEA (ablation) | SVEA · Walker, walk color_hard | episode return: 760 -> 877 | Helped |
| Add augmented-target bootstrapping (ablation) | SVEA with CNN · Walker, walk color_hard | episode return: 760 -> 605 | Hurt |
| Replace SVEA objective with mix-all objective (ablation) | SVEA with CNN · Walker, walk color_hard | episode return: 760 -> 599 | Hurt |
| Replace SVEA objective with ordinary Q objective while retaining strong augmentation (ablation) | SVEA with CNN · Walker, walk color_hard | episode return: 760 -> 569 | Hurt |
| Remove strong convolution augmentation (ablation) | SVEA with CNN · Walker, walk color_hard | episode return: 760 -> 498 | Hurt |
| Add strong convolution augmentation to DrQ (ablation) | DrQ with CNN · Walker, walk color_hard | episode return: 520 -> 569 | Mixed |
| Add strong convolution augmentation to DrQ with ViT (ablation) | DrQ with ViT · Walker, walk color_hard | episode return: 94 -> 470 | Helped |
| Add strong convolution augmentation to RAD (ablation) | RAD with CNN · Walker, walk color_hard | episode return: 400 -> 246 | Hurt |
| Use SVEA rather than DrQ with CNN | SAC with CNN · Robotic reach test | success rate: 0.60 -> 0.89 | Helped |
| Use SVEA rather than DrQ with CNN | SAC with CNN · Robotic mv.tgt. test | success rate: 0.69 -> 0.96 | Helped |
| Use SVEA rather than DrQ with CNN | SAC with CNN · Robotic push test | success rate: 0.26 -> 0.48 | Helped |
Closest papers by meaning
- Revisiting Data Augmentation in Deep Reinforcement Learning (ICLR 2024 · similarity 0.90)
- Zero Shot Generalization of Vision-Based RL Without Data Augmentation (ICML 2025 · similarity 0.89)
- MAD-TD: Model-Augmented Data stabilizes High Update Ratio RL (ICLR 2025 · similarity 0.88)
- NoisyRollout: Reinforcing Visual Reasoning with Data Augmentation (NeurIPS 2025 · similarity 0.87)
- Generalizing Consistency Policy to Visual RL with Prioritized Proximal Experience Regularization (NeurIPS 2024 · similarity 0.87)
- Reinforcement Learning with Euclidean Data Augmentation for State-Based Continuous Control (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.