Dongkkka/bgf_groot_30man

BGF GR00T SH5 tactile checkpoint — 300,000 steps

Inference-only package from completed Task000650 fine-tuning at 300,000 steps. It includes all model shards, index, model configuration, saved processor, normalization statistics and embodiment mapping required by GR00T inference. Optimizer/scheduler/RNG state is intentionally excluded.

Contract

• Robot: ffw_sh5_rev1; action order: left arm (7), right arm (7), left hand (20), right hand (20).
• Tactile: two 45D hands (90 raw taxels), saved normalized processor, finger_mlp to 32D latent.
• Inputs: three RGB cameras and language; action output is an absolute 40×54 chunk.
• Load with Gr00tPolicy(model_path=..., embodiment_tag="new_embodiment", use_tactile=True).

Required files

Keep model-*.safetensors, model.safetensors.index.json, config.json, processor_config.json, statistics.json, and embodiment_id.json together. Do not rename a shard to model.pt. SHA256SUMS covers the package files except itself and Hub .gitattributes.

See inference_manifest.json, Python modality config, CYCLO_INFERENCE.md, and QUALITY_AUDIT.md. The audit is a training-set offline replay diagnostic; this checkpoint is not approved for powered robot operation.

Python training configuration

The former configs/sh5.json has been converted to configs/ffw_sh5_rev1/ffw_sh5_rev1_config.py. The same folder includes the dataset array mapping, modality.json, and a training/inference guide.

This is a configuration-format migration: the 54 joint states/actions, three camera keys, task-description key, 40-step action horizon, and optional tactile settings are equivalent to the original JSON. Historical training source commits and run configurations in this package remain provenance records.

For further training, use a GR00T launcher that reads the Python exports MODALITY_CONFIG, EMBODIMENT_TAG, and TACTILE_CONFIG. This loader extension was added after the historical training commits; the converted file alone does not add that support to an older launcher. Pass:

--modality-config-path <downloaded_model_dir>/configs/ffw_sh5_rev1/ffw_sh5_rev1_config.py \
--use-tactile \
--letter-box-transform False

The Python file is for training configuration. Inference continues to load config.json, processor_config.json, statistics.json, and embodiment_id.json with the model shards. Those JSON files remain required; no parameter-name conversion or retraining is needed for this metadata update.

전처리 안내: 이 체크포인트의 processor_config.json → processor_kwargs.letter_box_transform은 false를 유지해야 합니다. 저장된 전처리는 패딩 없이 256×256으로 리사이즈하고, 학습 시 무작위 / 추론 시 중앙 230×230 영역을 잘라 다시 256×256으로 리사이즈합니다. 다른 회사 체크포인트의 레터박스 설정을 이 모델에 적용하지 마세요.