H-JEPA, DeepJEPA and RoboJEPA: what must a machine predict to plan?
TheTuringPost · x · 2026-10-21
The Turing Post explores JEPA's core idea — predict what matters without recreating every pixel — via three variants answering "what does a machine need to predict well enough to plan?": H-JEPA plans at distinct levels from local details to the big picture; DeepJEPA learns when a prediction warrants a second look (more compute); RoboJEPA scales up JEPA's predictor for better predictions and thus better actions. The author hopes a robot combining these ideas will change how it acts, with a full video walkthrough of each.
Related event: Breaking down three JEPA variants: what machines must predict to plan(2 posts)→
More from Embodied
- Basalt monocular-inertial VIO ported to WASM: SLAM now runs live in mobile browsers — rsasaki0109 · 2026-10-21
- DIY voice-enabled AI companion device with 3D-printed shell and pet avatar — shekitup · 2026-10-21
- RPL humanoid perceptive locomotion code released, accepted to CoRL 2026 — GuanyaShi · 2026-10-21
- RoboARC shows robots get no free retries—unlike coding agents that can just rerun tests — ManlingLi_ · 2026-10-21
- All 20 robot arms failed in overcast weather at demo day, sparking low-level control debate — chris_j_paxton · 2026-10-21
- 'They still fail the burrito test': the case for putting AI models in bodies — generativist · 2026-10-21