JEPA4R: Embodied JEPA World Models with 4D Point Planning
Trained only in a digital twin, JEPA4R runs zero-shot on a real robot arm with either a gripper or a dexterous hand.
The model uses point clouds from a single depth camera to predict and plan in latent space. A decoder makes its predictions visible as 4D point clouds, with reported mean point error within 1.1 cm over 200 steps. Under the cited benchmark protocols, it outperforms pixel-based JEPA models on four OGBench tasks. It also ranks first on MetaWorld and Adroit Relocate, scoring 98.4% and 88.0%, respectively.
HF Daily Papers' note
The model uses point clouds from a single depth camera to predict and plan in latent space. A decoder makes its predictions visible as 4D point clouds, with reported mean point error within 1.1 cm over 200 steps. Under the cited benchmark protocols, it outperforms pixel-based JEPA models on four OGBench tasks. It also ranks first on MetaWorld and Adroit Relocate, scoring 98.4% and 88.0%, respectively.
HF Daily Papers' note
score 5