FineCraft: final MetaFine Challenge policies

Model assets for FineCraft: Enhancing VLA Models for Fine-Grained Manipulation, the HITSZ Robotics entry that placed fourth in the MetaFine Challenge.

Paper · PDF · Code

This bundle contains only the five policies used for the final leaderboard results. The code is maintained in the FineCraft GitHub repository. The original native checkpoint formats and asset directory layout are preserved.

Task	Final policy	Scored revision	Official score*
T1	body r25 + cap r3, continuity v4	cf5b770c0608	0.796
T2	pi0.5 LoRA, step 29999	25cb20a04623	1.000
T3	pi0.5 stage4 coverage3000	cf5b770c0608	0.925
T4	pi0.5 LoRA, step 29999	25cb20a04623	0.987
T5	general/finishing/correction combination	cf5b770c0608	0.681

* Rounded to three decimals. Overall official score: 0.878. Official leaderboard.

Contents

The 79 inference assets total 23,490,243,495 bytes (23.49 GB). manifest.json records each file's size, SHA256, task and source revision. SHA256SUMS also covers the model card, manifest and component notices.

• t1_best/: paired body/cap action and perception weights, DINOv2 configuration, and the scored deployment_options.json.
• checkpoints/grasp_move_mug/29999/: T2 Orbax parameters and normalization.
• checkpoints/toggle_switch_table/coverage3000_20261019/: T3 parameters and normalization.
• checkpoints/put_blocks_into_boxes/29999/: T4 parameters and normalization.
• checkpoints/insert_letter/combo_20261019/: all seven T5 checkpoints.
• models/t5_combo/dinov2-large/config.json: T5 vision configuration.
• tokenizer/paligemma_tokenizer.model: the bundled tokenizer.

Small Orbax metadata/manifest files are required to read the parameter shards. Do not remove them based on file size or extension. Component checkpoint filenames retain the original development labels; these labels are not standalone official scores for the final combination.

Usage

Use the companion FineCraft inference code on Linux with Python 3.11 and an NVIDIA GPU. Point FINECRAFT_MODEL_DIR at this directory, then run:

python policy_server.py --port 8080

The input is two 512 x 512 RGB views, nine joint positions, a task ID and an instruction. The policy server returns seven arm commands and one gripper command through the official MetaFine HTTP interface.

The pi0.5 checkpoints use native Orbax/JAX parameter storage. T1/T5 use native PyTorch checkpoint payloads loaded by the companion model definitions. The bundle is intended for those loaders rather than a generic Transformers from_pretrained call.

The code's verification tool checks the complete bundle without deserializing the tensors. Training data, historical variants, optimizer runs, logs, videos and hidden evaluation seeds are outside this release.

Citation

If you use FineCraft in your research, please cite our technical report:

FineCraft: Enhancing VLA Models for Fine-Grained Manipulation (arXiv:2611.03163, 2026).

@misc{zhang2026finecraft,
  title         = {{FineCraft}: Enhancing {VLA} Models for Fine-Grained Manipulation},
  author        = {Shuaike Zhang and Ziheng Quan and Junhao Ran and
                   Muhe Ding and Shaokun Wang and Jianlong Wu},
  year          = {2026},
  eprint        = {2611.03163},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2611.03163}
}

Citation metadata: BibTeX · CITATION.cff.

Provenance and terms

Every inference asset is copied byte-for-byte from the scored submission and checked against its sealed SHA256 list. T2/T4 are traced to their original scored revision; the later scored integrated submission retained those assets.

See LICENSES.md and licenses/ for component notices and upstream references. No blanket relicensing of external pretrained weights or tokenizer is claimed.