AaronHuangWei/pi05-robocasa365-composite-seen-sft-100k

pi0.5 RoboCasa365 Composite-Seen SFT — 100k updates

Final supervised fine-tuning checkpoint for the 16 RoboCasa365 Composite-Seen tasks, initialized from the pi0.5 Human300 checkpoint at step 75,000. This release contains the completed 100,000-update SFT policy, before PPO training.

Training used the pi05_robocasa_finetune_target_composite_seen recipe, 500 human demonstrations per task, a global batch of 64, and 16 H100 80GB GPUs across two nodes. The checkpoint step is 99999 because training uses zero-based indices. The last logged training loss was 0.0507 at step 99,900; this is a training loss, not an evaluation success rate. No evaluation success-rate result is included.

The model page is public. Download access requires approval from the repository owner. Request access on this page, then authenticate with your Hugging Face account before downloading.

Files

• model.safetensors, config.json, robocasa365/norm_stats.json: BF16 PyTorch export for the pinned OpenPI/RLinf RoboCasa365 adapter, approximately 7.2 GB.
• jax/99999/: original completed Orbax checkpoint, including parameters, training state, and normalization assets for resuming SFT, approximately 44.7 GB.
• training_info.json, dataset_manifest.json: training and source provenance.
• file_manifest.json: sizes and SHA-256 checksums for the published files.

Download the PyTorch export

from huggingface_hub import snapshot_download

checkpoint = snapshot_download(
    repo_id="AaronHuangWei/pi05-robocasa365-composite-seen-sft-100k",
    token=True,
    ignore_patterns=["jax/**"],
)

Use this directory as PI05_SFT_CKPT with the RoboCasa365 OpenPI/RLinf adapter. Preserve its mean/std normalization using robocasa365/norm_stats.json. The PyTorch export is not a generic Transformers AutoModel checkpoint. Omit ignore_patterns to download the complete JAX training checkpoint as well.

Provenance

Training code: StructRL_RoboCasa365, commit fdf3561, with local checkpoint retention and interactive allocation resume support. OpenPI: robocasa-benchmark/openpi, commit 5a6beda9ff99da30b4e1b59320f6a32971d7c397. Base weights: robocasa/robocasa365_checkpoints, revision c484448aba1a9b60a04c9b0ca117241518ea69f3, pi05_pretrain_human300/multitask_learning/75000. See the JSON manifests for the fixed dataset revision and task/date paths.

The upstream checkpoint repository declares Apache-2.0 licensing. This model release preserves that license designation.