Research map / Reinforcement learning / Visual and representation RL
Capturing Visual Environment Structure Correlates with Control Performance
Jiahua Dong, Yunze Man, Pavel Tokmakov, Yu-Xiong Wang
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 |
|---|---|---|---|
| use full state-prediction proxy instead of Few-Shot/Action MSE/ImageNet/Segmentation/Depth proxies | visual backbone selection · avg over 4 sim environments | MMRV: 0.035 vs best baseline 0.068 | Improved |
| use state-prediction proxy instead of baselines | visual backbone selection · avg over 4 sim environments | Pearson r: 0.753 vs 0.347 (Few-Shot) | Improved |
| add state prediction auxiliary loss to policy fine-tuning | ViT-IN, Mocov3, MAE, CLIP, DINOv2 · MetaWorld | success rate: 0.683 -> 0.740 (ViT-IN); 0.765 -> 0.801 (CLIP) | Improved |
| use only a single state dimension instead of the full state vector (ablation) | state prediction proxy · 4 sim environments | MMRV: e.g. 0.037 -> 0.153 (material only, Metaworld) | Hurt |
| use full-episode Few-Shot proxy instead of 5-episode version (ablation) | Few-Shot proxy · MetaWorld | Pearson r: 0.650 -> 0.885 (MetaWorld); 0.624 -> 0.348 (SimplerEnv G) | Mixed |
| switch diffusion policy to MLP policy (ablation) | state prediction proxy · policy algorithm robustness | correlation with success | No clear effect |
| use state prediction proxy instead of Few-Shot in compute cost | ViT-IN selection · MetaWorld | time: 113 min -> 12 min | Improved |
| finetune backbones rather than freeze (ablation) | visual backbones · MetaWorld | success rate ranking | Mixed |
Closest papers by meaning
- Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers (NeurIPS 2024 · similarity 0.88)
- Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Datasets (ICLR 2025 · similarity 0.88)
- DynaFLIP: Rethinking Robotics Perception via Tri-Modal-Dynamics Guided Representation (NeurIPS 2026 · similarity 0.88)
- SyncWorld: Visual Calibration Enables World Models as Zero-Shot Simulators (NeurIPS 2026 · similarity 0.88)
- DMC-VB: A Benchmark for Representation Learning for Control with Visual Distractors (NeurIPS 2024 · similarity 0.87)
- Video Prediction Policy: A Generalist Robot Policy with Predictive Visual Representations (ICML 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.
