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
• Code: zju3dv/INTACT-JEPA
• Project page: 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:

# 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:

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.