H-JEPA world model lifts maze task success rate to 73%

A global research team including the NYU AMI Lab, INRIA Paris and Brown University has published H-JEPA, a hierarchical world model, on an online archive, according to aitimes.com. In the AntMaze ant-robot maze escape environment, the three-level model raised the success rate from 18% for a single-level model to 73%.

H-JEPA separates short-term and long-term planning into different levels of abstraction. The levels are trained end-to-end, with each level predicting future states in its own latent space. Level 1 predicts short-term changes over 5 frames, while the higher levels use 10-frame and 20-frame scales.

A normalization mechanism called SIGReg stabilizes the inter-level training process. Planning then proceeds top-down: the highest level sets the general direction, and the lower levels turn it into concrete actions.

In DROID trials, an Inverse Dynamics loss improved trajectory fidelity, though the researchers evaluated offline trajectory reproduction only. They said H-JEPA is likely applicable to robot navigation, logistics automation and object manipulation, and they plan to connect higher-level representations to language so robots can follow natural language instructions.

Background

Yann LeCun took part in the research as a member of NYU AMI Lab. JEPA was a technology he developed during his time at Meta, and H-JEPA is a core world model technology at his startup AMI.