LeCun co-authored Temporal Straightening: straightening latent paths boosts gradient-based planning success rates
ylecun · x · 2026-10-26
A new paper co-authored by Yann LeCun, "Temporal Straightening for Latent Planning," tackles a core weakness of world models: latent trajectories from pretrained visual encoders are highly curved, so Euclidean distance in latent space misrepresents progress toward a goal and gradient-based planners struggle.
Key points:
- Inspired by perceptual straightening in human vision, the method adds a curvature regularizer during training to straighten latent trajectories, jointly learning a JEPA encoder and predictor;
- Straightened embeddings make Euclidean distance a better proxy for geodesic distance, smoothing the planning objective;
- Empirically, gradient-based planning becomes more stable and achieves significantly higher success rates across goal-reaching tasks, avoiding costly search methods like CEM/MPPPI;
- Code is open-sourced (agenticlearning.ai/temporal-straightening); the paper appears at ICML 2026.
For robotics, it's a promising direction: improve planning by changing the representation rather than rewriting RL algorithms.
More from Embodied
- OmniEgoCap Turns Any Egocentric Device into a Full-Body 3D Mocap System — Scobleizer · 2026-10-26
- Tesla FSD approved in Latvia, the 10th European country — mitchdeg · 2026-10-26
- On-device VLM goes production: GenieX runs Liquid AI's LFM-VL live at PyTorch Conference — PyTorch · 2026-10-26
- One long Claude Code session builds a full Android TV launcher with a 1-second-latency live camera — aristofeles · 2026-10-26
- Humanoid robots' DRAM demand may 20x, but it's still just 0.04% of 2026 supply — zephyr_z9 · 2026-10-26
- Microduck team publishes deep dives: relocalizes within 2cm using a 64-pixel depth sensor — mishig25 · 2026-10-26