Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2603.12231 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866908882151931904 |
|---|---|
| author | Wang, Ying Bounou, Oumayma Zhou, Gaoyue Balestriero, Randall Rudner, Tim G. J. LeCun, Yann Ren, Mengye |
| author_facet | Wang, Ying Bounou, Oumayma Zhou, Gaoyue Balestriero, Randall Rudner, Tim G. J. LeCun, Yann Ren, Mengye |
| contents | Learning good representations is essential for latent planning with world models. While pretrained visual encoders produce strong semantic visual features, they are not tailored to planning and contain information irrelevant -- or even detrimental -- to planning. Inspired by the perceptual straightening hypothesis in human visual processing, we introduce temporal straightening to improve representation learning for latent planning. Using a curvature regularizer that encourages locally straightened latent trajectories, we jointly learn an encoder and a predictor. We show that reducing curvature this way makes the Euclidean distance in latent space a better proxy for the geodesic distance and improves the conditioning of the planning objective. We demonstrate empirically that temporal straightening makes gradient-based planning more stable and yields significantly higher success rates across a suite of goal-reaching tasks. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2603_12231 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Temporal Straightening for Latent Planning Wang, Ying Bounou, Oumayma Zhou, Gaoyue Balestriero, Randall Rudner, Tim G. J. LeCun, Yann Ren, Mengye Machine Learning Learning good representations is essential for latent planning with world models. While pretrained visual encoders produce strong semantic visual features, they are not tailored to planning and contain information irrelevant -- or even detrimental -- to planning. Inspired by the perceptual straightening hypothesis in human visual processing, we introduce temporal straightening to improve representation learning for latent planning. Using a curvature regularizer that encourages locally straightened latent trajectories, we jointly learn an encoder and a predictor. We show that reducing curvature this way makes the Euclidean distance in latent space a better proxy for the geodesic distance and improves the conditioning of the planning objective. We demonstrate empirically that temporal straightening makes gradient-based planning more stable and yields significantly higher success rates across a suite of goal-reaching tasks. |
| title | Temporal Straightening for Latent Planning |
| topic | Machine Learning |
| url | https://arxiv.org/abs/2603.12231 |