Latent Dynamics · 2026-10-06
H-JEPA: End-to-End Learning of Hierarchical World Models for Visual Planning
Wancong Zhang, Basile Terver, Michael Rabbat, Yann LeCun, Randall Balestriero
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
H-JEPA is an end-to-end hierarchical, action-conditioned JEPA world model. It trains multiple JEPA levels where each level predicts farther ahead in its own learned latent space. Top-down planning uses higher levels to predict subgoals for lower levels, discarding fast unpredictable detail and retaining slow task-relevant state. The authors report improved hierarchical planning over a flat JEPA on simulated AntMaze and other environments, and extension to real-robot videos with inverse-dynamics supervision.
Why it matters
Long-horizon latent planning can suffer from compounded prediction error and a mismatch between a single shared latent space used for both fast dynamics and goal matching. The authors argue that hierarchical latent spaces plus temporal decomposition can provide more informative costs at higher abstraction levels, enabling lower compute planning and better success on long-horizon visual tasks.
Method
- Train a stack of action-conditioned JEPAs end-to-end: each level ℓ predicts future latent states using its own encoder/predictor and a temporal stride/window from the level below.
- Use top-down hierarchical gradient-based planning: the top level optimizes macro-actions to reach the goal in its latent space; predicted intermediate states become subgoals for lower levels.
- Optionally add inverse-dynamics supervision when extending to DROID real-robot videos, so the hierarchy is trained with action-label information.
Limitation
The authors report that hierarchical planning over three levels can perform poorly on Push-T: “the three-level model performs poorly even with maximum compute,” likely due to short episodes limiting usable training data (three-level Push-T sees only 58% of LeWM’s transitions per epoch).
Abstract (from arXiv)
Long-horizon planning with latent world models requires reasoning across timescales and levels of abstraction. Existing task-agnostic JEPA world models predict and plan at a single timescale or with multiple horizons in one shared latent space. We introduce H-JEPA, an end-to-end recipe for training a hierarchy of action-conditioned JEPAs in which each level predicts farther ahead in its own learned latent space. Planning proceeds top-down: the top level optimizes progress toward the goal, and each level's predictions become subgoals for the planner below it. When factors in the data evolve at separated timescales, higher levels discard fast, unpredictable detail and retain slower task-relevant state. Across four simulated navigation and manipulation environments, hierarchical planning improves over a flat JEPA; on Visual AntMaze, a three-level hierarchy raises success from 18% to 73% using less planner compute. Ablations attribute these gains to both temporal decomposition and higher-level goal representations. With inverse-dynamics supervision, the approach extends to diverse real-robot videos from DROID, where hierarchy improves offline planning fidelity at lower planner compute.
Related papers
- Beyond Action Prediction: Learning Search-Free Planning in Latent Space
- LRC-JEPA: Disentangling Dynamics and Residual Context for Efficient World Models
- LeWAM: End-to-End World and Action Modeling with JEPAs
- JEPA-TTT: Persistent Test-Time Training of Latent World Models for Planning under Dynamics Shifts
- The Planning Limits of Latent World Models
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-05 · Indexed by Pub-AI: 2026-10-06 UTC
H-JEPA: End-to-End Learning of Hierarchical World Models for Visual Planning
Preprint posted: 2026-10-05 · Original record ↗
- arxiv
- 2610.06805
Explicit source versions (1)
- v1 · 2026-10-05
- Preprint available
Source date: 2026-10-05 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Long-horizon planning with latent world models requires reasoning across timescales and levels of abstraction. Existing task-agnostic JEPA world models predict and plan at a single timescale or with multiple horizons in one shared latent space. We introduce H-JEPA, an end-to-end recipe for training a hierarchy of action-conditioned JEPAs in which each level predicts farther ahead in its own learned latent space. Planning proceeds top-down: the top level optimizes progress toward the goal, and each level's predictions become subgoals for the planner below it. When factors in the data evolve at separated timescales, higher levels discard fast, unpredictable detail and retain slower task-relevant state. Across four simulated navigation and manipulation environments, hierarchical planning improves over a flat JEPA; on Visual AntMaze, a three-level hierarchy raises success from 18% to 73% using less planner compute. Ablations attribute these gains to both temporal decomposition and higher-level goal representations. With inverse-dynamics supervision, the approach extends to diverse real-robot videos from DROID, where hierarchy improves offline planning fidelity at lower planner compute.