Fast-JEPA

Fast-JEPA is a latent world-model planner for visual goals. It builds a reference action plan from the current observation and goal image, then compares complete candidates by predicting their endpoints directly. The authors report 96.50% mean success across four simulation tasks.

View full model description

Low-latency, goal-image-conditioned action-sequence planning in simulation.

Papers & reports (1)

Fast-JEPA: Advancing the Success–Latency Frontier for Latent World Models

Paper available

arXiv: 2611.25927
Authors & affiliations (7)

Author order: Original order

  1. Yao Zhang

    Wuhan University · Zhongguancun Academy · DeepCybo

  2. Bin Yu

    Harbin Institute of Technology · Zhongguancun Academy

  3. Shijie Lian

    Huazhong University of Science and Technology · Zhongguancun Academy

  4. Haishan Liu

    Nanjing University · Zhongguancun Academy

  5. Changti Wu

    East China Normal University · Zhongguancun Academy

  6. Yu Wucorresponding_author

    Wuhan University

  7. Kai Chencorresponding_author

    Zhongguancun Academy · Zhongguancun Institute of Artificial Intelligence · DeepCybo

First published
2026-11-20
Latest paper revision
v1
Venue
Not provided

Covered model versions

Paper ↗PDF ↗

How to access and use

Training / fine-tuning code

Access
Not publicly available

SourceVerified on · 2026-11-22

Release information

SourceVerified on · 2026-11-22

Checks before deployment

The paper reports simulations. The authors have yet to provide the implementation, checkpoints and DynaGoal evaluation assets. Available weights or code do not establish plug-and-play robot compatibility. Check the required configuration, data and interfaces in the source documentation.

Architecture and inputs / outputs

Model type
World simulation
Architecture
Transformer
Backbone
ViT-Tiny image encoder
Input
Goal image / RGB images

Scope and training

Functional purpose

Task planning / World prediction

Output

Actions / Reasoning / plans

Prediction output

192-dimensional latent state

Training

Imitation learning

Evaluation results

No verified independent benchmark results are recorded for this version or the selected weights.

Publisher evaluation · Reported; not independently reproduced

Native four-task simulation · Native four-task simulation ↗

Mean success rate: 96.5 %

NVIDIA RTX PRO 6000 Blackwell Server Edition

100 episodes per task, three independently trained models. Warmed batch-one complete planning.

Fast-JEPA · Native four-task simulation · PushT / Reacher / Cube / TwoRoom

Publisher evaluation · Reported; not independently reproduced

Native four-task simulation · Native four-task simulation ↗

Mean complete planning latency: 10.88 ms

NVIDIA RTX PRO 6000 Blackwell Server Edition

100 episodes per task, three independently trained models. Warmed batch-one complete planning.

Fast-JEPA · Native four-task simulation · PushT / Reacher / Cube / TwoRoom

Publisher evaluation · Reported; not independently reproduced

DynaGoal · Shift ↗

Mean success rate: 73.83 %

NVIDIA RTX PRO 6000 Blackwell Server Edition

100 cases per condition, three original-task models. Measured planning delay determines availability of each plan in simulator ticks.

Fast-JEPA · Shift · PushT / Reacher / Cube / TwoRoom

Publisher evaluation · Reported; not independently reproduced

DynaGoal · Shift ↗

Mean complete planning latency: 13.42 ms

NVIDIA RTX PRO 6000 Blackwell Server Edition

100 cases per condition, three original-task models. Measured planning delay determines availability of each plan in simulator ticks.

Fast-JEPA · Shift · PushT / Reacher / Cube / TwoRoom

Publisher evaluation · Reported; not independently reproduced

DynaGoal · Deadline ↗

Mean success rate: 87.67 %

NVIDIA RTX PRO 6000 Blackwell Server Edition

100 cases per condition, three original-task models. The deadline counts actual planning time and action execution time.

Fast-JEPA · Deadline · PushT / Reacher / Cube / TwoRoom

Publisher evaluation · Reported; not independently reproduced

DynaGoal · Deadline ↗

Mean complete planning latency: 13.75 ms

NVIDIA RTX PRO 6000 Blackwell Server Edition

100 cases per condition, three original-task models. The deadline counts actual planning time and action execution time.

Fast-JEPA · Deadline · PushT / Reacher / Cube / TwoRoom

Full parameters and verification notes (model card)

The planner first builds a reference action sequence using intermediate outcome predictions. It then uses the cross-entropy method (CEM) to refine candidates, scoring their predicted terminal latent states relative to the goal image. It executes a prefix and replans from the next observation.

Table 1 reports 96.50% ± 0.66 percentage points average success and 10.88 ± 0.04 ms mean complete planning time. The local INTACT comparison used 26.19 ± 0.22 ms (2.41× ratio). The four tasks were PushT, Reacher, Cube and TwoRoom. The authors used 100 episodes per task and three independently trained models, with warm batch-one timing on an NVIDIA RTX PRO 6000 Blackwell Server Edition GPU.

DynaGoal tests changing goals or geometry (Shift), execution deadlines (Deadline), and controlled delays (Delay). The authors reported 73.83% mean success for Shift and 87.67% for Deadline without benchmark-specific fine-tuning.

Submit a source or correction