pravsels/pi05-bin-pack-reward-recap-positive-only-from-base

pi0.5 Bin Pack — Reward Recap (Positive Only, from Base)

Fine-tuned pi0.5 checkpoint for coffee capsule bin packing, trained with positive-only advantage conditioning (reward recap) starting from pi0.5 base weights (no task-specific pre-training).

Experiment

• Objective: Test whether positive-only advantage conditioning works when training from pi0.5 base weights directly (no task-specific pre-training).
• Weight init: weights/pi05_base/params (pi0.5 base weights).
• Advantage mode: positive_only — human demos are trained with prompt "pack coffee capsules into the cardboard bin container. Advantage: positive", policy-collected frames are dropped.
• Target steps: 100,000

Config

• Config name: pi05_bin_pack_coffee_capsules_reward_recap_positive_only_from_base
• 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

Dataset

9 LeRobot datasets (1 base + 8 dAgger rounds):

• villekuosmanen/bin_pick_pack_coffee_capsules
• villekuosmanen/dAgger_bin_pick_pack_coffee_capsules_1.0.0
• villekuosmanen/dAgger_bin_pick_pack_coffee_capsules_1.1.0
• villekuosmanen/dAgger_bin_pick_pack_coffee_capsules_1.2.0
• villekuosmanen/dAgger_bin_pick_pack_coffee_capsules_1.3.1
• villekuosmanen/dAgger_bin_pick_pack_coffee_capsules_1.4.0
• villekuosmanen/dAgger_bin_pick_pack_coffee_capsules_1.5.0
• villekuosmanen/dAgger_bin_pick_pack_coffee_capsules_1.5.1
• villekuosmanen/dAgger_bin_pick_pack_coffee_capsules_1.7.0

Loss Progression

Step	Loss
1,100	0.0425
25,000	0.0097
50,000	0.0066

Checkpoint Hashes

Verify integrity with tar cf - -C checkpoints/<step> params | sha256sum.

Step	Loss	SHA-256
25,000	0.0097	fe1e6b97b7dafa1ea6e74a35f698df798ba2b739ecef54c38a810beffb404e75
50,000	0.0066	27d2ceb579a0cdd7b34a06c2c6c52b3b8bcdbbdcd81b0ed2826ee62daaaa6603
73,000	0.0053	a43e1edbb84b6050588f7d1bec5158ba2c5b4649613586a6105cf9a8318ceb00

Repo Structure

assets/                      # Norm stats for inference
checkpoints/<step>/params/   # Model weights (params only)
README.md                    # This file
TRAINING_LOG.md              # Training log

W&B

• Training dashboard

Usage

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

config = get_config("pi05_bin_pack_coffee_capsules_reward_recap_positive_only_from_base")
server = PolicyServer(config, checkpoint_path="checkpoints/<step>/params")