pi05_pytorch_robocasa_open_drawer_contact

π₀.₅ PyTorch — RoboCasa OpenDrawerContact

Fine-tuned π₀.₅ (OpenPI PyTorch) checkpoint on RoboCasa OpenDrawerContact.

• Config: pi05_pytorch_robocasa_open_drawer_contact
• Run: open_drawer_contact_clip_bs32
• Base weights: /data/disk_14t/wg/ckpts/openpi/pi05_base_torch (π₀.₅ PyTorch base)
• Dataset: open_drawer_contact (OpenDrawer, 100_demos, horizon 730)
• Train steps: 20,000 | batch size: 32 | save interval: 10,000
• Model: π₀.₅, PaliGemma gemma_2b + action expert gemma_300m, bfloat16, action_dim=32, action_horizon=50, max_token_len=200

License: Apache-2.0 (same as OpenPI). The Gemma/PaliGemma backbone remains subject to the Gemma Terms of Use.

Files

Recommended checkpoint is step 20000.

open_drawer_contact_clip_bs32/
  10000/                  # intermediate
    model.safetensors
    metadata.pt
    assets/open_drawer_contact/norm_stats.json
  20000/                  # final / recommended
    model.safetensors
    metadata.pt
    assets/open_drawer_contact/norm_stats.json

optimizer.pt is not uploaded (resume-only, ~12.5G per step).

Load

After huggingface-cli download Tsaochengyu/pi05_pytorch_robocasa_open_drawer_contact --local-dir <dir>:

from openpi.training import config as _config
from openpi.policies import policy_config as _policy_config

train_cfg = _config.get_config("pi05_pytorch_robocasa_open_drawer_contact")
ckpt = "<dir>/open_drawer_contact_clip_bs32/20000"
policy = _policy_config.create_trained_policy(train_cfg, ckpt)

The loader expects model.safetensors plus assets/ in that step directory.