pravsels/pi05-so101-stacking-rings

pi0.5 SO101 Stacking Rings

Fine-tuned pi0.5 checkpoint for the SO101 stacking-rings task (6D joint-space actions with delta joints + absolute gripper).

Note: Training was in progress when step 5000 was published. Later steps may be added to this repo under checkpoints/<step>/params/.

Experiment

• Objective: Replication run — verify pi0.5 fine-tuning on lorenzouttini/so101_stacking_rings.
• Weight init: weights/pi05_base/params (pi0.5 base weights).
• Published step: 5,000 (of 50,000 planned).
• Loss at step 5,000: 0.0252

Config

• Config name: pi05_so101_stacking_rings
• Model: pi0.5 (pi05=True, action_horizon=30)
• Batch size: 32
• Learning rate: 2.5e-5 cosine decay (1k warmup, decay to 2.5e-6)
• EMA decay: 0.999
• Delta actions: mask [T,T,T,T,T,F] (5 joints delta, gripper absolute)
• Norm: quantile normalization (pi0.5 default) + per-timestep action norm
• Default prompt: stack the rings

Dataset

• lorenzouttini/so101_stacking_rings — 101 episodes, ~34k frames

Checkpoint Hashes

Verify integrity with:

cd checkpoints/<step> && find params -type f | sort | xargs sha256sum | sha256sum

Step	Loss	SHA-256
5,000	0.0252	3e772c819c5e0233b939e5f739f4de74b1ca8224e4fcf9499d59a9bf603cdb7c

W&B

• Training dashboard

Repo Structure

assets/                         # Norm stats, valid_indices.txt
checkpoints/5000/params/        # Model weights (params only)
README.md                       # This file
TRAINING_LOG.md                 # Training log

Usage

from openpi.training.config import get_config
from openpi.serving.policy_server import PolicyServer

config = get_config("pi05_so101_stacking_rings")
server = PolicyServer(config, checkpoint_path="checkpoints/5000/params")