Research map / Probabilistic and scientific modeling / Molecules and proteins
Temporal Straightening for Latent Planning
Ying Wang, Oumayma Bounou, Gaoyue Zhou, Randall Balestriero, Tim G. J. Rudner, Yann LeCun, Mengye Ren
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 curvature regularizer to a projector on frozen DINOv2 patch features (ablation) | DINOv2 patch + proj 14x14x8 · PointMaze UMaze open-loop GD | success rate (%): 44.00 -> 94.00 | Improved |
| add curvature regularizer to ResNet trained from scratch with spatial features (ablation) | ResNet 14x14x8 · Wall open-loop GD | success rate (%): 1.33 -> 84.67 | Improved |
| add curvature regularizer to ResNet with global features (ablation) | ResNet 1x384 · PushT MPC | success rate (%): 70.00 -> 83.33 | Improved |
| add curvature regularizer to projector with global features (ablation) | DINOv2 patch + proj 1x384 · PushT | success rate (%): 2.00 -> 2.00 open-loop; 11.33 -> 8.67 MPC | No clear effect |
| train encoder/projector with JEPA loss only (no explicit straightening) instead of frozen DINOv2 patch (ablation) | DINOv2 patch + proj · Wall open-loop GD | success rate (%): 52.67 -> 80.00 | Improved |
| add smoothness or temporal contrastive regularizer instead of curvature (ablation) | DINOv2 + spatial projector · PushT open-loop GD | success rate (%) | No clear effect |
| use learnable aggregation head for curvature loss instead of patch, mean or flatten cosine (ablation) | DINOv2 + spatial projector · Wall, UMaze, Medium, PushT open-loop | success rate (%) | Improved |
| increase channel dimension of projected DINOv2 features from 8 to 128 (ablation) | DINOv2 + spatial projector · Wall, mazes, PushT open-loop | success rate | Hurt |
| collapse patch features into a single global vector (ablation) | DINOv2 + projector · all four environments | success rate | Hurt |
| use CEM instead of GD with straightened encoder (ablation) | DINOv2 patch + proj + Lcurv · PointMaze UMaze open-loop | success rate (%): 94.00 -> 94.00 | No clear effect |
| add global goal cost on aggregated features with weight 0.1 for long-horizon planning (ablation) | Straightened models, MPC · PointMaze Medium, PushT long horizon | success rate (%): PushT ResNet 33.33 -> 36.00; Proj 24.00 -> 33.33 | Improved |
| add curvature regularizer to ResNet for long-horizon planning (ablation) | ResNet · PointMaze Medium 50-step MPC | success rate (%): 48.00 -> 98.67 | Improved |
Closest papers by meaning
- Predictive but Not Plannable: RC-aux for Latent World Models (NeurIPS 2026 · similarity 0.86)
- Latent Plan Transformer for Trajectory Abstraction: Planning as Latent Space Inference (NeurIPS 2024 · similarity 0.86)
- World-aware Planning Narratives Enhance Large Vision-Language Model Planner (NeurIPS 2025 · similarity 0.86)
- Sparse Planning in Visual World Models via Cost Gradients (NeurIPS 2026 · similarity 0.85)
- PcLast: Discovering Plannable Continuous Latent States (ICML 2024 · similarity 0.85)
- State-Covering Trajectory Stitching for Diffusion Planners (NeurIPS 2025 · similarity 0.85)
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.
