Research map / Reinforcement learning / Visual and representation RL
PQDA:Policy-Aligned Q-Consistency Meets Decoupled Augmentation for Generalizable Visual RL
Yun Zhou, Yuqiang Wu, Chunyu Tan
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 |
|---|---|---|---|
| add policy-aligned Q-consistency (PAQC) to SAC (ablation) | SAC · CVDMC color-easy video-easy | average return: 548 -> 606 | Improved |
| add foreground-background decoupled augmentation (FBDA) on top of PAQC (ablation) | SAC w/ PAQC · CVDMC color-easy video-easy | average return: 606 -> 822 | Improved |
| replace original Q-regularization with PAQC (ablation) | SMG · CVDMC color-easy video-easy | average return: 763 -> 791 | Improved |
| replace original Q-regularization with PAQC (ablation) | SGQN · CVDMC color-easy video-easy | average return: 691 -> 753 | Improved |
| replace PAQC with original Q-consistency (ablation) | PQDA · CVDMC color-easy video-easy | average return: 822 -> 797 | Hurt |
| use GMM mask instead of ground-truth mask (ablation) | PQDA · CVDMC color-easy video-easy | average return: 816 -> 822 | No clear effect |
| replace SGQN saliency mask with GMM mask (ablation) | SGQN · CVDMC color-easy video-easy | average return and training time: 691 -> 803; 13h -> 10.5h | Improved |
| replace SMG mask with GMM mask (ablation) | SMG · CVDMC color-easy video-easy | average return and training time: 763 -> 808; 17h -> 13.5h | Improved |
| use FBDA instead of background-only augmentation (ablation) | PQDA · CVDMC color-easy video-easy | average return: 772 -> 822 | Improved |
| use FBDA instead of random overlay augmentation (ablation) | PQDA · CVDMC color-easy video-easy | average return: 761 -> 822 | Improved |
| use PQDA instead of SMG | visual RL generalization · DMControl video-hard, cheetah run | return: 303±46 (SMG) vs 185±42 | Hurt |
| use PQDA instead of SAC and SGQN regarding training cost | visual RL · DMControl with color and video distractions | reward and training time: 217 -> 810 reward; 5.5h -> 6.5h | Improved |
Closest papers by meaning
- Generalizing Consistency Policy to Visual RL with Prioritized Proximal Experience Regularization (NeurIPS 2024 · similarity 0.89)
- What Effects the Generalization in Visual Reinforcement Learning: Policy Consistency with Truncated Return Prediction (AAAI 2024 · similarity 0.89)
- Zero Shot Generalization of Vision-Based RL Without Data Augmentation (ICML 2025 · similarity 0.88)
- A Simple Framework for Generalization in Visual RL under Dynamic Scene Perturbations (NeurIPS 2024 · similarity 0.88)
- Focus On What Matters: Separated Models For Visual-Based RL Generalization (NeurIPS 2024 · similarity 0.88)
- VIGOR: Zero-Shot Visual Generalization via Latent-Space Consistency in Model-Based Reinforcement Learning (NeurIPS 2026 · similarity 0.88)
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.
