pi05-so101-lora-100demos

LoRA fine-tune of pi05_base on jakegonz/pick-and-place-red-block-100demos.

Step	Folder
5000	step_5000/
10000	step_10000/
15000	step_15000/

Training config

• Base: gs://openpi-assets/checkpoints/pi05_base/params
• Train config: pi05_so101_lora (PaliGemma 2B LoRA + Gemma 300M LoRA action expert)
• Action horizon: 10 (≈0.33 s @ 30 Hz)
• Batch size: 32
• LR schedule: cosine, peak 5e-5, warmup 500 steps, decay over 10k
• Optimizer: AdamW with grad clip 1.0

Inference inputs (per step)

observation.state              : float32 (6,)  joint pos (5 arm + 1 gripper, 0=open/100=closed)
observation.images.camera1     : uint8        wrist camera   → maps to `left_wrist_0_rgb`
observation.images.camera2     : uint8        overhead camera → maps to `base_0_rgb`
prompt                         : "Pick up the red block and place it"

Loading

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

cfg = _config.get_config("pi05_so101_lora")
policy = policy_config.create_trained_policy(cfg, "step_15000")

Each step folder contains params/, assets/, and _CHECKPOINT_METADATA — the minimum required by create_trained_policy. The train_state/ optimizer state is not included.