Research map / Reinforcement learning / Visual and representation RL
WorldGym: World Model as An Environment for Policy Evaluation
Julian Hector Quevedo, Ansh Kumar Sharma, Yixiang Sun, Varad Suryavanshi, Percy Liang, Sherry Yang
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 |
|---|---|---|---|
| train world model on larger Bridge V2 data instead of Bridge V1 subset (ablation) | WorldGym world model · Bridge validation videos | LPIPS: 0.131 -> 0.073 | Improved |
| set diffusion prediction horizon equal to policy chunk size (h=4 vs h=1) (ablation) | WorldGym inference · 40-frame rollouts on A100 | inference time (s): 93 -> 33 | Improved |
| evaluate policies in WorldGym instead of real world | RT-1-X, Octo, OpenVLA · OpenVLA Bridge eval suite | mean success rate (%): 70.6 -> 67.4 (OpenVLA) | No clear effect |
| add random OOD distractor objects to initial frames | OpenVLA · OpenVLA Bridge eval suite (WorldGym) | mean success rate (%): 67.4 -> 39.4 | Hurt |
| add random OOD distractor objects to initial frames | Octo · OpenVLA Bridge eval suite (WorldGym) | mean success rate (%): 23.82 -> 4.1 | Hurt |
| add random OOD distractor objects to initial frames | RT-1-X · OpenVLA Bridge eval suite (WorldGym) | mean success rate (%): 15.5 -> 7.6 | Hurt |
| recolor carrot red so it differs from orange target | OpenVLA · Bridge OOD: pick up orange | grasp correctness | Improved |
| add laptop displaying a carrot image as distractor | OpenVLA · Bridge OOD: pick up carrot | wrong-object grasp rate: 15% of trials grab the screen carrot | Hurt |
| use checkpoints with more training steps for UniPi and DexVLA policies | UniPi, DexVLA · Bridge V2 policies in WorldGym | WorldGym success rate | Improved |
Closest papers by meaning
- World Action Verifier: Self-Improving World Models via Forward-Inverse Asymmetry (NeurIPS 2026 · similarity 0.90)
- Action Images: End-to-End Policy Learning via Multiview Video Generation (NeurIPS 2026 · similarity 0.89)
- SyncWorld: Visual Calibration Enables World Models as Zero-Shot Simulators (NeurIPS 2026 · similarity 0.89)
- From Forecasting to Planning: Policy World Model for Collaborative State-Action Prediction (NeurIPS 2025 · similarity 0.89)
- EA-WM: Event-Aware Generative World Model with Structured Kinematic-to-Visual Action Fields (NeurIPS 2026 · similarity 0.88)
- WorldModelBench: Judging Video Generation Models As World Models (NeurIPS 2025 · 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.
