Seonglae ChoJoint Embedding Predictive Architecture
JEPA Types
Making representations for latent planning where state-change trajectories in representation space are less curved makes planning much easier. Existing pretrained visual encoders have highly curved latent trajectories, meaning that Euclidean distance in latent space poorly reflects actual reachability difficulty or geodesic distance, and gradient-based planning also fails. Therefore, the authors propose temporal straightening: Reducing Curvature
arxiv.org
https://arxiv.org/pdf/2603.12231
Temporal Straightening for Latent Planning
Inspired by the perceptual straightening hypothesis in human vision, this paper introduce temporal straightening to JEPA world models to facilitate latent planning
arxiv.org
https://arxiv.org/pdf/2603.12231
Temporal Straightening for Latent Planning | Agentic Learning AI Lab
Inspired by the perceptual straightening hypothesis in human vision, we introduce temporal straightening to improve representation learning for latent planning.
https://agenticlearning.ai/temporal-straightening/
