luobaigu/Snap2Policy-Checkpoints

Snap2Policy checkpoints

Archived Pi0.5 robot policy weights for three manipulation scenes.

Scene	Original	Meridian	Lyra	Marble
Drawer	Available	Available	Available	Available
Sink / bowl	Available	Available	Available	—
Takeaway bag	Available	Available	Available	Available

Each method uses the same directory layout: models/<scene>/<method>/ contains model.pt, policy_config.yaml, training_config.json, and checkpoint_metadata.json. The archive contains 11 full policy state dictionaries. No completed Marble sink checkpoint was available when this archive was prepared.

Download

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="luobaigu/Snap2Policy-Checkpoints",
    allow_patterns=[
        "models/drawer/meridian/*",
        "shared_assets/*",
        "manifest.json",
    ],
)

Replace the scene and method to select another policy. Scene directory names are drawer, sink, and takeaway_bag; method directory names are lowercase.

Policy interface and assets

These are full PyTorch state dictionaries from the project's RLinf/openpi PyTorch training workflow, rather than a Transformers from_pretrained package. Loading requires the compatible Pi0.5 model implementation and base model assets. Configure actor.model.model_path to the base model directory and actor.model.openpi.assets_dir to this archive's absolute shared_assets path, then load the chosen model.pt using the project's policy loader.

• Input camera: left_wrist_0_rgb only.
• State dimension: 8; raw environment action dimension: 8; padded model action dimension: 32.
• Action chunk length: 10; RTC training maximum delay: 4.
• Shared normalization asset: all400_state8_abs_eef8; mean/std normalization, with quantile normalization disabled.
• Shared normalization statistics are included under shared_assets/.

manifest.json and per-checkpoint metadata record the actual checkpoint step, byte count, and SHA-256 checksum. training_config.json preserves the recorded training settings. This archive contains inference weights and configuration; optimizer state, training datasets, and base model assets are separate.

No benchmark scores are asserted by this archive. The method names identify the source variants, not a performance ranking.