Three JEPA variants, one question: how well must a machine predict to plan?
TheTuringPost · x · 2026-10-21
TheTuringPost explores JEPA's core idea—predicting what matters without reconstructing every pixel—and how three variants tackle the question of what a machine needs to predict well enough to plan actions:
- H-JEPA: planning at distinct levels, from fine local details to the big picture
- DeepJEPA: learning when a prediction needs a second look (more computation)
- RoboJEPA: scaling up JEPA's predictor for better predictions and thus better actions
The author hopes to see a robot combining these ideas and how that would change its behavior. Full breakdown in the linked video.
Related event: Breaking down three JEPA variants: what machines must predict to plan(2 posts)→
More from Embodied
- NVIDIA DGX Spark Founders Edition lands, creator preps local AI series — LinusEkenstam · 2026-10-21
- Shenzhen robotics founder meets Scobleizer: hardware-software fusion as the US-China bridge — Scobleizer · 2026-10-21
- DeepMind robotics head joins Berkeley, Boston Dynamics panel on humanoid motion & perception — VLA vs agentic coding debate — chris_j_paxton · 2026-10-21
- YC F26's Dreamscale Labs: Four College Dropouts Building Cloud Inference for Robotics — ycombinator · 2026-10-21
- Origami Robotics Teases Infant-Home Dexterity Demo, Full Reveal Tomorrow — ybisk · 2026-10-21
- Dream Labs exits stealth with $90M seed, ships Dream-1 world action model API at $3/hour — km · 2026-10-21