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

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

Fine-tuned pi0.5 checkpoint for coffee capsule bin packing, trained with positive-only advantage conditioning (reward recap).

Experiment

• Objective: Test whether positive-only advantage conditioning improves bin-pack policy when fine-tuning from a task-trained checkpoint.
• Weight init: Resumed from pi05-bin-pack-single-dataset checkpoint (step 29999).
• 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
• 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
3,100	0.0148
25,000	0.0074
50,000	0.0058

Checkpoint Hashes

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

Step	Loss	SHA-256
25,000	0.0074	23558b103ffeccb94dead23db3adf0c9119f38338d7a7ddc171db579a83bf6b1
50,000	0.0058	7b705f798619f1ecf4d5e8773896684ac735844265bcd0649cdd9d6dc18b5207
80,000	0.0050	f0cbcbf79a6072e33696e83d20540b9b9367d0c48bb1ace0d5d48b8d17981ccc

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