Temporal Straightening
Summary
Temporal Straightening is a representation-geometry regularizer for action-conditioned latent world models. It aligns consecutive latent velocity vectors so observed trajectories become locally straighter, with the aim of making latent Euclidean distance a more useful goal cost for gradient-based planning.
Interface
Given consecutive latent states,
v_t=z_{t+1}-z_t, \qquad \mathcal L_{\mathrm{curv}} = 1- rac{v_t^\top v_{t+1}}{\lVert v_t\rVert_2\lVert v_{t+1}\rVert_2}.The published system combines this loss with stop-gradient next-latent prediction in a JEPA world model and then differentiates a latent goal cost through candidate action/control-input rollouts.
observations + actions/control inputs
-> latent state and action embeddings
-> predictive latent dynamics
-> local trajectory-curvature regularization
-> candidate rollout
-> latent goal cost
-> GD or MPC plannerRole In The Wiki
Temporal Straightening is the explicit geometry branch of the local JEPA/world-model line:
- LeWorldModel reports temporal straightening as an emergent training property without an explicit curvature objective.
- Temporal Straightening turns that property into a direct loss and connects it to planner conditioning under a linear-dynamics theory scope.
- AdaJEPA uses Temporal-Straightening world models as frozen bases and adapts them from deployment transitions.
- SkyJEPA measures temporal straightness of physical-control latent rollouts but does not use the explicit loss as its main training objective.
The transferable time-series lesson is not “all latent trajectories should be straight.” It is that representation geometry should be tested against downstream planning utility. Irregular sampling, meaningful regime changes, asymmetric dynamics, and typed interventions require time normalization and possibly directional costs.
Evidence Boundary
The accepted ICML 2026 paper reports gains on Wall, PointMaze-UMaze, PointMaze-Medium, and PushT. The strongest claims are visual 2D goal-reaching results under the paper’s own GD/MPC protocol. No numeric multivariate time-series experiment, physical-robot deployment, checkpoint reproduction, or independent replication was verified during ingest.
The user-provided X post is third-party commentary by Haiyu Wu, not an author announcement. Its “36% versus 0%” long-horizon shorthand does not match the camera-ready table: 36% is the straightened ResNet MPC result with a combined cost, while the DINO-WM row reports 3.33% open-loop and 27.33% MPC.