---
language:
- en
viewer: false
tags:
- world-model
- jepa
- decision-supervision
- robotics
task_categories:
- robotics
---

# D-JEPA decision supervision

**[Project website](https://nebulis-lab.com/D-JEPA)** ·
**[Paper](https://arxiv.org/abs/2609.24749)** ·
**[Model repository](https://huggingface.co/Shuaijun/D-JEPA)** ·
**[Decision-supervision dataset](https://huggingface.co/datasets/Shuaijun/D-JEPA-Dataset)** ·
**[Code](https://github.com/NEBULIS-Lab/D-JEPA)**

## Download

### Release versions

| Version | Contents | Availability |
|---|---|---|
| v1 | PushT, Reacher, Granular, unseen-shape PushObj and visual-condition supervision | Available below |
| v1.5 | Compact RoboTwin manipulation and autonomous-driving supervision; no repeated v1 data | Incremental download below |
| v2.0 | Physical-robot supervision | Planned for a later release |

### v1.5 — RoboTwin and autonomous driving

Download **[D-JEPA-supervision-v1.5.zip](https://huggingface.co/datasets/Shuaijun/D-JEPA-Dataset/resolve/main/D-JEPA-supervision-v1.5.zip?download=true)**
(16,946,070 bytes; approximately 16.2 MiB). Verify it with
[manifest-v1.5.json](./manifest-v1.5.json).

- **RoboTwin:** compact roller-grasping supervision, 27 starts and 99 executed
  candidates; candidate development, gate development and validation remain separate.
  This lightweight subset omits videos, parameter sweeps and repeated recordings.
- **Driving:** 200 fitting, 50 calibration and 200 test scenes, each with 32
  trajectories, predictive features and simulator-scored outcomes (14,400 candidates).

The archive contains only these additions, not v1 data or physical-robot data.
See [release notes](./v1.5/README.md) and [field definitions](./v1.5/SCHEMA.md).

```bash
hf download Shuaijun/D-JEPA-Dataset D-JEPA-supervision-v1.5.zip manifest-v1.5.json \
  --repo-type dataset --local-dir .
python -m zipfile -e D-JEPA-supervision-v1.5.zip data
```

### v1 — Original supervision package

Download **[D-JEPA-supervision-v1.zip](https://huggingface.co/datasets/Shuaijun/D-JEPA-Dataset/resolve/main/D-JEPA-supervision-v1.zip?download=true)**
(487,257,636 bytes; approximately 487 MB) and verify its SHA256 against the
[manifest](./manifest.json). The archive has one top-level `D-JEPA-supervision-v1/` folder.
It includes candidate inputs, decision supervision, training/calibration
identities and recorded results for the completed paper experiments.

```bash
hf download Shuaijun/D-JEPA-Dataset D-JEPA-supervision-v1.zip manifest.json \
  --repo-type dataset --local-dir .
python -m zipfile -e D-JEPA-supervision-v1.zip data
```

After extraction, start with `data/D-JEPA-supervision-v1/SCHEMA.md` and
`POPULATIONS.json`. Each split has `inputs.npz`, `metadata.json`, and, where
observed candidate supervision exists, `labels.npz`.

```python
import numpy as np

with np.load("data/D-JEPA-supervision-v1/pushobj/train/inputs.npz",
             allow_pickle=False) as inputs:
    candidate_ids = inputs["candidate_ids"]
    features = inputs["features"]
```

## Browser preview and loading

This release is a **downloadable ZIP of NPZ arrays**, not a single tabular
dataset. Automatic Dataset Viewer processing is disabled because it cannot
interpret this mixed archive as training rows. This does not restrict access:
download the ZIP above or open the **Files and versions** tab.

Use `hf download` followed by `numpy.load(..., allow_pickle=False)` as shown
above, rather than `datasets.load_dataset("Shuaijun/D-JEPA-Dataset")`.
The actual task-specific splits and schemas are defined in the archive's
`POPULATIONS.json` and `SCHEMA.md`; an automatically inferred Hub
`default/train` split does not describe those populations.

## Data and protocol

The archive is designed for lightweight learning and cached-decision evaluation.
It is not a complete simulator installation or an exhaustive raw-video dataset.

Important data semantics:

- Inference arrays and outcome labels are separate files.
- Granular has **sparse** supervision: unlabeled candidates are unknown, not failures.
- PushT's fitting pool has module-specific fit/calibration assignments.
- The independent PushT confirmation has 256 starts, distinct from development.
- Corruption-trained PushT has 50 independent confirmation starts repeated under
  seven known conditions; do not treat its 350 condition rows as independent starts.
- Some tasks provide training features and separate recorded formal results rather
  than all formal candidate labels. Only actual observed labels are released.

The detailed schema and population inventory are inside the ZIP. No personal
data are intentionally included; observations and decision outcomes originate
from the recorded simulated tasks. The authors must finalize the dataset license
and applicable upstream resource terms before external publication.