AMI Labs Universities jointly announce the H-JEPA hierarchical world model, with maze-task success rate reaching 73.3%

AT A GLANCE

H-JEPA used a hierarchical world model to improve maze-planning success, but the results still require verification against the original materials.

Article

Yang Likun 4 Years-old concept comes to fruition: H-JEPA AntMaze success rate rises from 18%to 73%Beating AI news reports that AMI Labs founded by Yang Likun, together with multiple universities, has announced H-JEPA, a world model capable of hierarchically predicting the future and planning actions. The high level handles long-term goals, while the low level handles specific actions. The paper, code, and pretrained weights have all been made public.

H-JEPA extends the approach of JEPA. It first converts images into abstract environmental features, then combines them with actions to predict future states, without having to generate video frame by frame. However, long-term planning and immediate control require different information. For example, when a quadruped robot navigates a maze, route selection mainly depends on position and corridors, whereas taking steps requires body posture. A single-layer model uses the same features for both tasks, making farther-ahead planning increasingly difficult.

H-JEPA combines multiple JEPA into different planning levels. The high level predicts farther into the future, passes intermediate goals to lower levels, and progressively refines them into specific actions. Unlike HWM which previously shared the same features across all levels, H-JEPA allows each level to learn its own environmental features. In maze experiments, the high level retained positional information while gradually ignoring short-term details such as leg posture. Yang Likun proposed a similar concept as early as 2022 years ago, and the team has now provided an end-to-end trainable implementation.

In the Visual AntMaze simulated maze, the three-level H-JEPA achieved a success rate of 73.3%, compared with 18.0%for the single-layer LeWorldModel, while requiring less planning computation. However, more planning levels are not necessarily better. In the simulated object-pushing Push-T task, the three-level model actually underperformed the two-level model. The team also conducted offline tests using the real-robot video dataset DROID, and its predicted action trajectories were closer to expert demonstrations than those of the single-layer model.

Original article link https://m.theblockbeats.info/flash/371155

Key points

AMI Labs jointly announced H- with multiple universities JEPA。
H-JEPA can hierarchically predict the future and plan actions, with the high level responsible for long-term goals and the low level responsible for specific actions.
The paper, code, and pretrained weights have all been made public.
In the Visual AntMaze simulated maze, the three-level H-JEPA had a success rate of 73.3%, while the single-layer LeWorldModel achieved 18.0%。
In the Push-T task, the three-level model underperformed the two-level model.
The team used the real-robot video dataset DROID for offline testing, and its predicted action trajectories were closer to expert demonstrations than those of the single-layer model.
Yang Likun proposed a similar concept in 2022 year.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

H-JEPA breaks planning into different levels: the high level determines farther-ahead goals, while lower levels progressively refine them into specific actions. It enables different levels to learn different environmental features, allowing the high level to focus on position and corridors and the low level to focus on short-term details such as body posture.

Why it matters to readers

The maze experiments described in the materials show that H-JEPA had a higher success rate than the single-layer LeWorldModel and required less planning computation. This suggests that hierarchical prediction may help robots handle longer-horizon action planning.

can be understood as a robot model that “first decides where to go, then decides how to take each step.” It does not need to generate video frame by frame; instead, it predicts abstract future environmental states.

Risks and unknowns

The performance data comes from the experiments described in the materials; the original publication links for the paper, code, and pretrained weights, as well as their accessibility, still require verification.
The evaluation setup and baseline configuration for Visual AntMaze are not described in detail in the materials.
The materials do not provide results from online deployment on real robots or failure cases.
The three-layer model performs worse than the two-layer model on the Push-T task, indicating that more planning levels are not necessarily better.

AMI Labs Universities jointly announce the H-JEPA hierarchical world model, with maze-task success rate reaching 73.3%

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center