---
license: mit
tags:
  - robotics
  - world-model
  - jepa
  - imitation-learning
  - goal-conditioned-control
  - arxiv:2607.26056
---

# INTACT Checkpoints

Official checkpoints for **INTACT: Isomorphic Intent-to-Action Learning for
Search-Free World Models**.

[Paper](https://arxiv.org/abs/2607.26056) ·
[Code](https://github.com/zju3dv/INTACT-JEPA) ·
[Project page](https://zju3dv.github.io/INTACT-JEPA/)

## Audited results

| Setting | Inference | PushT | Cube | Reacher | TwoRoom | Macro SR |
|---|---|---:|---:|---:|---:|---:|
| **Task-specific E1 INTACT** | **Direct, zero search** | 87.44 +/- 1.26 | 100.00 +/- 0.00 | 97.33 +/- 0.00 | 97.67 +/- 1.20 | **95.61 +/- 0.59** |
| **Task-specific E1 INTACT** | **Guarded A 128x3** | 91.56 +/- 0.69 | 99.67 +/- 0.33 | 97.56 +/- 0.51 | 97.56 +/- 1.26 | **96.58 +/- 0.44** |
| **Shared-encoder E5 INTACT** | **Direct, zero search** | 86.11 +/- 0.96 | 100.00 +/- 0.00 | 97.22 +/- 0.84 | 81.56 +/- 3.67 | **91.22 +/- 0.51** |
| [History-free E1 ablation](https://huggingface.co/INTACT-JEPA/INTACT/tree/main/INTACT-no-previous-action) | Direct, zero search | 85.33 +/- 0.58 | 99.44 +/- 0.38 | 93.44 +/- 0.38 | 98.78 +/- 0.69 | **94.25 +/- 0.08** |

All values use the corrected Official evaluation implementation. Standard
INTACT uses strictly causal pre-`t` action history at a declared continuation
point and appends every policy-executed action exactly once. The history-free
ablation removes the previous-action channel from both training and evaluation.
Guarded A is an optional local verifier around the Direct plan; it evaluates
384 sampled candidates rather than the 9,000 used by CEM 300x30.
Machine-readable headline values are frozen in
[`INTACT/AUDITED_RESULTS_20260914.json`](INTACT/AUDITED_RESULTS_20260914.json)
and
[`INTACT-unified/AUDITED_RESULTS_20260914.json`](INTACT-unified/AUDITED_RESULTS_20260914.json).
Model manifests retain archive identities and hashes; use the audited result
files for project-level reporting.

## Download

```bash
# Standard INTACT: three training seeds, four task shards per archive.
hf download INTACT-JEPA/INTACT --include "INTACT/*" --local-dir checkpoints-release

# History-free INTACT.
hf download INTACT-JEPA/INTACT --include "INTACT-no-previous-action/*" \
  --local-dir checkpoints-release

# Complete unified E5 comparison matrix.
hf download INTACT-JEPA/INTACT --include "INTACT-unified/*" \
  --local-dir checkpoints-release
```

Each directory contains a model-specific README, auditable manifests, and
SHA-256 checksums. The unified directory additionally contains the complete
shared-encoder comparison matrix.

## Unified E5 comparison

| Controlled cell | Native interface | Official macro SR |
|---|---|---:|
| LeWM | CEM 300x30 | 66.17 |
| Inverse only | Direct | 68.19 |
| Goal intent only | Direct | 84.92 |
| **Goal-displacement INTACT** | **Direct** | **91.22** |

All cells use the same E5 shared-encoder protocol. Task-level statistics,
checkpoint hashes, and runtime fingerprints are provided
inside [`INTACT-unified`](https://huggingface.co/INTACT-JEPA/INTACT/tree/main/INTACT-unified).
The immutable original layout also remains available at the frozen
[`paper-e5-goal-v1`](https://huggingface.co/INTACT-JEPA/INTACT/tree/paper-e5-goal-v1)
tag. The original E1 root layout is preserved at
[`intact-goal-e1-single-p77-v1`](https://huggingface.co/INTACT-JEPA/INTACT/tree/intact-goal-e1-single-p77-v1).

## Citation

```bibtex
@article{sun2026intact,
  title   = {INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models},
  author  = {Sun, Junhan and Zhao, Hao and Zhang, Guofeng},
  journal = {arXiv preprint arXiv:2607.26056},
  year    = {2026}
}
```