Researchers including Yann LeCun unveil H-JEPA, a hierarchical world model; simulated maze success rises from 18% to 73% (report)

What happened

According to AI Times, a joint team from NYU's AMI Lab, INRIA Paris, Brown University and others unveiled the hierarchical world model H-JEPA on Oct. 5 (local time), with Yann LeCun participating in the research. Existing world models handle short-term detailed actions and long-term goals in a single latent space, which reduces prediction accuracy. H-JEPA processes time horizons and levels of abstraction separately, layer by layer. In simulated maze navigation experiments, it reportedly raised the success rate from 18% for a single-level model to 73% and also reduced the compute required for planning. The results are based on simulated environments, and the report does not say whether they were validated on real robots.

Provider claims

Researchers: separating hierarchy levels improved long-horizon planning success and reduced planning compute (as relayed in the report).

Why it matters

This applies the JEPA-style world models LeCun has long advocated to long-horizon robot planning. It can be seen as one approach to long-task planning, a current weakness of VLA and world models.

Sources

• Press 얀 르쿤의 월드 모델 'H-제파', 단·장기 계획 분리로 로봇 경로 계획 성공률 73% 달성 AI Times (Korea)