pi0.5 Bin Pack — Single Dataset Baseline
Fine-tuned pi0.5 checkpoint for coffee capsule bin packing, trained on a single dataset of ~200 teleoperated episodes. This serves as the base checkpoint for the reward recap experiments.
Config
- Config name:
pi05_bin_pack_coffee_capsules_delta_single_dataset - Model: pi0.5 (
pi05=True,action_horizon=50) - Batch size: 36
- Learning rate: 5e-5 cosine decay (10k warmup)
- Optimizer: AdamW (gradient clip norm 1.0)
- EMA decay: 0.999
- Delta actions: enabled
- Weight init:
weights/pi05_base/params(pi0.5 base weights) - Training steps: 30,000
Dataset
villekuosmanen/bin_pick_pack_coffee_capsules(~200 teleoperated episodes)
Checkpoint Hash
Verify integrity with tar cf - -C checkpoints/29999 params | sha256sum.
| Step | SHA-256 |
|---|---|
| 29,999 | bb051b5a3ee10adae7ee5313102fd7157e49d77a12a3b9a48e0688617108f9b0 |
Downstream
This checkpoint is the weight init for the reward recap experiments:
Repo Structure
assets/ # Norm stats + valid indices for inference
checkpoints/29999/params/ # Model weights (params only)
README.md # This file
Usage
from openpi.training.config import get_config
from openpi.serving.policy_server import PolicyServer
config = get_config("pi05_bin_pack_coffee_capsules_delta_single_dataset")
server = PolicyServer(config, checkpoint_path="checkpoints/29999/params")