Research map / Reinforcement learning / Visual and representation RL
Persistent Robot World Models: Stabilizing Multi-Step Rollouts via Reinforcement Learning
Jai Bardhan, Patrik Drozdík, Josef Sivic, Vladimir Petrik
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 RL post-training on own rollouts to Ctrl-World | PersistWorld · DROID val, external cameras | PSNR: 23.02 -> 24.42 | Improved |
| add RL post-training on own rollouts to Ctrl-World | PersistWorld · DROID val, external cameras | LPIPS: 0.081 -> 0.070 | Improved |
| add RL post-training on own rollouts to Ctrl-World | PersistWorld · DROID val, wrist camera | SSIM: 0.62 -> 0.67 | Improved |
| add RL post-training on own rollouts to Ctrl-World | PersistWorld · DROID val, wrist camera | PSNR: 17.80 -> 19.39 | Improved |
| add RL post-training to Ctrl-World, object-only masked | PersistWorld · DROID val, external camera object-only | LPIPS: 0.025 -> 0.021 | Improved |
| add RL post-training to Ctrl-World, robot-only masked | PersistWorld · DROID val, robot-only | PSNR: 17.62 -> 19.25 | Improved |
| replace Self-Forcing style regression with the RL contrastive objective (ablation) | PersistWorld · DROID val, 10-step rollout, external | PSNR: 25.35 -> 25.52 | Improved |
| replace Self-Forcing style regression with the RL contrastive objective (ablation) | PersistWorld · DROID val, 10-step rollout, wrist | PSNR: 20.17 -> 20.40 | Improved |
| replace Self-Forcing style regression with the RL contrastive objective (ablation) | PersistWorld · DROID val, 10-step rollout, wrist | LPIPS: 0.338 -> 0.338 | No clear effect |
| lower learning rate from 3e-4 to 1e-4 (ablation) | PersistWorld · DROID val, 10-step rollout, external | PSNR: 25.02 -> 25.52 | Improved |
| use PersistWorld instead of Ctrl-World for policy evaluation | PersistWorld · Policy evaluation (3 tasks, 3 policies) | MMRV: 0.053 -> 0.006 | Improved |
| use PersistWorld-generated data to improve pi0 | PersistWorld · Real-robot block picking | task progression: 0.558 -> 0.667 | Improved |
Closest papers by meaning
- RoboAlign-R1: Distilled Multimodal Reward Alignment for Robot Video World Models (NeurIPS 2026 · similarity 0.92)
- SyncWorld: Visual Calibration Enables World Models as Zero-Shot Simulators (NeurIPS 2026 · similarity 0.90)
- World Action Verifier: Self-Improving World Models via Forward-Inverse Asymmetry (NeurIPS 2026 · similarity 0.89)
- Learning 3D Persistent Embodied World Models (NeurIPS 2025 · similarity 0.89)
- EA-WM: Event-Aware Generative World Model with Structured Kinematic-to-Visual Action Fields (NeurIPS 2026 · similarity 0.89)
- Learning Visual Feature-Based World Models via Residual Latent Action (NeurIPS 2026 · similarity 0.89)
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.
