File size: 3,154 Bytes
ae86829
 
2727a48
 
 
 
 
 
ae86829
2727a48
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0430df6
2727a48
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
---
license: mit
tags:
  - robotics
  - world-model
  - jepa
  - imitation-learning
  - goal-conditioned-control
---

# INTACT Paper Checkpoints

This repository hosts the paper checkpoint matrix for **INTACT: Isomorphic
Intent-to-Action Learning for Search-Free World Models**.

- Paper: [arXiv:2607.26056](https://arxiv.org/abs/2607.26056)
- Code: [zju3dv/INTACT-JEPA](https://github.com/zju3dv/INTACT-JEPA)
- Project page: [zju3dv.github.io/INTACT-JEPA](https://zju3dv.github.io/INTACT-JEPA/)
- Frozen model revision: `paper-e5-goal-v1`

## Contents

The release contains six controlled shared-encoder E5 cells, three training
seeds (`0`, `42`, and `3072`), and four task-specific shards per seed (PushT,
Cube, Reacher, and TwoRoom): `6 x 3 x 4 = 72` checkpoints.

| Cell | Archive prefix | Native interface | Official macro SR |
|---|---|---|---:|
| LeWM | `paper-e5-lewm-*` | CEM 300x30 | 66.17 |
| Inverse only | `paper-e5-inverse-only-*` | Direct | 56.67 |
| Waypoint intent only | `paper-e5-waypoint-intent-only-*` | Direct | 74.22 |
| Goal intent only | `paper-e5-goal-intent-only-*` | Direct | 80.61 |
| Waypoint INTACT | `paper-e5-waypoint-intact-*` | Direct | 76.42 |
| Goal-displacement INTACT | `intact-goal-e5-*` | Direct | **89.39** |

Success rates are percentages averaged over three independently trained seeds.
Each checkpoint value averages 100 episodes for evaluation seeds `0`, `1`, and
`42`. See `PAPER_E5_MATRIX_MANIFEST.json` for task-level means, sample standard
deviations, checkpoint paths, and SHA-256 hashes. Official LeWM and CLEAR-LeWM
results are separate protocols and must not be mixed.

## Recommended download

From the accompanying INTACT source release:

```bash
# One cell, all training seeds.
bash scripts/download_paper_checkpoints.sh waypoint_intact all

# Complete 72-checkpoint matrix.
bash scripts/download_paper_checkpoints.sh matrix all
```

The wrapper pins `paper-e5-goal-v1`, checks the pinned `SHA256SUMS` record,
verifies every downloaded archive, extracts it into `STABLEWM_HOME`, and then
verifies every `.pt` shard against the checked-in manifest.

For a direct Hub download:

```bash
hf download INTACT-JEPA/INTACT \
  intact-goal-e5-seed3072.tar.gz SHA256SUMS \
  --revision paper-e5-goal-v1 --local-dir checkpoints-release

cd checkpoints-release
sha256sum -c SHA256SUMS --ignore-missing
```

## Runtime compatibility

These paper checkpoints use the legacy five-slot Actor checkpoint grammar and
must be evaluated with the frozen `paper_runtime/` bundled in the source
release. The clean training runtime uses the zero-free four-slot Fig. 1 grammar;
the Actor parameter shapes are intentionally incompatible. Use the provided
paper evaluation wrappers instead of loading these weights through the clean
root runtime.

## Integrity

- `SHA256SUMS` authenticates all 18 archives.
- `PAPER_E5_MATRIX_MANIFEST.json` records all 72 extracted shard hashes.
- `RUNTIME_SHA256SUMS` records the frozen evaluator runtime fingerprint.

## License and citation

The release is provided under the MIT License. Please cite the INTACT paper;
the complete BibTeX entry is maintained in the source repository.