Liukaikai/pi0_doll_low_mem

Pi0 LoRA fine-tuned weights on the Doll dataset

Pi0 LoRA fine-tuned weights on the Doll dataset (Franka + joint-space actions). Deploy-only: params + assets only (no train_state).

Model

• Base: OpenPI Pi0 (LoRA: gemma_2b_lora + gemma_300m_lora)
• Data: nokaikai/doll_lerobot_v2 (7 joints + 1 gripper, absolute joint targets)
• Use: Load for inference; input image + current 8-dim state + prompt -> output 8-dim action (7 joints + 1 gripper)

Usage (openpi)

# Install openpi and download this repo, or use repo_id directly
pip install huggingface_hub
huggingface-cli download nokaikai/pi0_doll_low_mem --local-dir ./pi0_doll_low_mem

# Point checkpoint to the step directory (e.g. 4999)
python pi0_deploy.py \
  --checkpoint_dir ./pi0_doll_low_mem/doll_lora_20260222_184432/4999 \
  --config_name pi0_doll_low_mem \
  --prompt "Pick up the doll"

Loading from code (after clone or snapshot_download):

from huggingface_hub import snapshot_download
path = snapshot_download(repo_id="nokaikai/pi0_doll_low_mem")
checkpoint_dir = f"{path}/doll_lora_20260222_184432/4999"
# Then use openpi create_trained_policy(config, checkpoint_dir)

Layout

doll_lora_20260222_184432/
  4999/
    params/          # Model parameters (JAX/Orbax)
    assets/          # norm_stats, etc.
    _CHECKPOINT_METADATA

License & Credits

Weights trained with OpenPI and Doll data; for research/personal use. Doll dataset: nokaikai/doll_lerobot_v2 on Hugging Face.