README.md

---
license: apache-2.0
base_model: lerobot/VLA-JEPA-Pretrain
pipeline_tag: robotics
library_name: coreai
tags:
- coreai
- core-ai
- coreai-fabric
- aimodel
- coreml
- apple
- apple-silicon
- on-device
- robotics
- vla
- vision-language-action
- action
---

> **Canonical:** [`kevinqz/VLA-JEPA-Pretrain-CoreAI`](https://huggingface.co/kevinqz/VLA-JEPA-Pretrain-CoreAI) — source of truth. 

# VLA-JEPA Pretrain (fabric)

> ⚠️ **Robot policy — needs a matching robot to actuate.** This is
> [lerobot/VLA-JEPA-Pretrain](https://huggingface.co/lerobot/VLA-JEPA-Pretrain) converted to an
> Apple Core AI `.aimodel`. Its output is a **normalized action chunk** for
> **droid_pretrain**. Run it on any other robot, or with mismatched calibration /
> normalization stats, and it emits floats that *look* valid but **actuate
> garbage** — the most dangerous silent-failure mode. It is a **base checkpoint**:
> fine-tune on your robot's data before expecting useful behavior.

An Apple Core AI conversion of
[lerobot/VLA-JEPA-Pretrain](https://huggingface.co/lerobot/VLA-JEPA-Pretrain) — a **robot policy**
that maps images + proprioceptive state (+ a language instruction, when the model
uses one) to a continuous action chunk (flow_matching sampler). Produced by
[coreai-fabric](https://github.com/kevinqz/coreai-fabric/blob/main/recipes/vla-jepa-pretrain.yaml) and indexed by
[coreai-catalog](https://github.com/kevinqz/coreai-catalog).

## Model facts

| Field | Value |
|---|---|
| Parameters | 3B |
| Architecture | transformer |
| Capabilities | vision-language-action, robotics |
| Embodiment | droid_pretrain |
| Sampling | flow_matching (4-step) |
| Quantization / precision | none / bfloat16 |
| On-disk size | 3.2 GB |
| Asset kind | split-export policy (qwen_context, action_denoise_step) + norm_stats |
| assetVersion | 2.0 |


## Use it — this needs host code you supply

A policy is **not** a chat model: there is no stock high-level Swift runtime for
it. The bundle is the split-export shape — an `encode` graph (run once per
observation) + a `denoise_step` graph (the host drives it `num_steps` times) +
`norm_stats.json` (un-normalization). **You supply the host loop** (the N-step
sampler + un-normalization) in Swift. Recommended integration: keep LeRobot's
Python `RobotClient` for the servos/cameras/calibration, and run inference
on-device — see the `io_contract` in the catalog for the exact tensors.

```bash
pip install coreai-catalog && coreai-catalog install vla-jepa-pretrain
```

## Requirements

- **Deployment: macOS 27.0+ / iOS 27.0+, Xcode 27+.** The asset serializes with `minimum_os v27`,
  so the on-device Swift runtime requires macOS/iOS 27+. A Mac on macOS 26 can
  convert and inspect it but not run it on-device.
- Apple Silicon. A matching robot to actuate (see the banner).

## Verification (action parity)

- **Gate A (structure): passed** — the bundle's layout + metadata were
  validated on real hardware; the graphs load.
- **Gate B — action_parity: 100.0% min chunk-cosine** vs the fp16 reference over 8 frames of `recorded episodes`, 4-step, fixed-noise (measured on Apple Silicon).
- This certifies the export is **numerically faithful to the source policy** — it
  does **NOT** certify real-world task success, embodiment transfer, or
  closed-loop stability. Closed-loop success-rate eval (LIBERO/ManiSkill):
  **not_run** (a separate future gate). Reproduce with `coreai-fabric verify`.

## Provenance

| Field | Value |
|---|---|
| Base model | [lerobot/VLA-JEPA-Pretrain](https://huggingface.co/lerobot/VLA-JEPA-Pretrain) @ `e946c3e5b538d760f4b4ff239d1b1c12090c041d` |
| Converted by | `models/vla_jepa/export.py` (version not reported) |
| Recipe | [vla-jepa-pretrain](https://github.com/kevinqz/coreai-fabric/blob/main/recipes/vla-jepa-pretrain.yaml) (recipe_source: fabric) |
| Precision / quantization | bfloat16 / none |
| Conversion date | 2026-07-07 |

Machine-readable, in this repo:
[`parity-report.json`](./parity-report.json) (gate results) ·
[`reproduce-manifest.json`](./reproduce-manifest.json) · [`LICENSE`](./LICENSE)
(upstream terms).

## License and attribution

Weights licensed **apache-2.0** — see the bundled `LICENSE`. This artifact is a **converted derivative** of the base policy: its
weights were converted to Apple Core AI format. The conversion itself is
community work.

## Links

- **Base model:** [lerobot/VLA-JEPA-Pretrain](https://huggingface.co/lerobot/VLA-JEPA-Pretrain)
- **Reproduce:** [recipe `vla-jepa-pretrain`](https://github.com/kevinqz/coreai-fabric/blob/main/recipes/vla-jepa-pretrain.yaml)
- **Index:** [coreai-catalog](https://github.com/kevinqz/coreai-catalog) — the
  neutral registry tying upstream ↔ this asset ↔ mirror together
- [HF Collection](https://huggingface.co/collections/kevinqz/coreai-apple-on-device-6a4879f21c7e1a87c99bcf5a)


## The on-device Core AI ecosystem

- [coreai-fabric](https://github.com/kevinqz/coreai-fabric) — the reproducible
  recipe → `.aimodel` pipeline that produced this asset.
- [coreai-catalog](https://github.com/kevinqz/coreai-catalog) — the index of Core
  AI models across the community, with provenance and integration snippets.
- [apple/coreai-models](https://github.com/apple/coreai-models) — Apple's official
  exporters and runtimes.
- [LeRobot](https://huggingface.co/lerobot) — the upstream robotics ecosystem.

## Not affiliated with Apple

Community conversion. Not produced, hosted, or endorsed by Apple. Apple and Core
AI are trademarks of Apple Inc., used here only to describe the target
runtime/format. This is an independent community conversion.