JEPA4R: point-cloud JEPA world model transfers zero-shot to real robot arms
AIGeeksGroup · hf · 2026-11-17
JEPA4R is an embodied JEPA world model that observes single-depth-camera point clouds, predicts and plans in latent space ("4D point planning"), and acts via a policy head on frozen latents.
Key results:
- A decoder on the frozen world model visualizes its imagination as 4D point clouds, with mean point error within 1.1 cm over 200 steps.
- Beats pixel JEPA models on OGBench Cube-Single/Double/Triple and Scene; ranks first on MetaWorld and Adroit Relocate at 98.4% and 88.0%, above runner-up R3M (96.0% / 65.3%).
- Trained only in a digital twin, it runs zero-shot on a real arm with both a gripper and a dexterous hand.
More from Embodied
- Generalist AI's Andy Barry to Present Foundation Models in Robotics at MIT on Nov 19 — HanGuo97 · 2026-11-17
- NVIDIA open-sources USDCraft-10K: 10K articulated OpenUSD assets for robot training — chris_j_paxton · 2026-11-17
- Hub open-sources HubEgo-100K: 44.7TB egocentric video dataset for robots — rohanpaul_ai · 2026-11-17
- Figure cofounder vs analyst: Chinese humanoids do gymnastics, but who actually ships? — chris_j_paxton · 2026-11-17
- Figure CEO Brett Adcock joins Peter Diamandis' Moonshots podcast on humanoid robots — PeterDiamandis · 2026-11-17
- Dev uses Claude Code to reverse-engineer a cheap AI badge into an open-source Claude voice assistant — ConsequenceThen217 · 2026-11-17