pravsels/pi05-bin-pack-reward-recap-mixed-from-base

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

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

Experiment

• Objective: Test whether mixed positive/negative 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: mixed — human demos are trained with prompt "pack coffee capsules into the cardboard bin container. Advantage: positive", policy-collected frames with "... Advantage: negative".
• Target steps: 100,000 (stopped at 74,200)

Config

• Config name: pi05_bin_pack_coffee_capsules_reward_recap_mixed_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
0	0.3473
25,000	0.0125
50,000	~0.0075
74,000	~0.0070

Note: High initial loss (0.35) is expected — mixed mode introduces negative demonstrations the base model hasn't seen. Loss dropped rapidly in the first few thousand steps, then converged to ~0.007 by 74k.

Checkpoint Hashes

Verify integrity with find params -type f | sort | xargs cat | sha256sum.

Step	Loss	SHA-256
25,000	0.0125	9e6d903b70a0159d6fb9979570556b031650f2e733e9b8a30c1d17b08f3307c2
50,000	~0.0075	f347d098e046f63ef65aa9c0c7a5614e0735667f1d03a5f8fb893e43698079c9
74,000	~0.0070	4fd1de8b341df95595b7691443e524638c80a0f1eb58c09d717a33741482d70e

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