donghao-zhou/PackLab-Bench

📌 Benchmark Summary

PackLab-Bench is a standardized benchmark for evaluating closed-loop robotic bin packing policies. It contains 60 test cases organized into easy, medium, and hard difficulty levels.

The benchmark is designed for evaluating whether a model can select objects and predict executable placement actions from multimodal packing states.

🗂️ File Structure

The released benchmark is organized as follows:

PackLab-Bench/
├── README.md
├── easy/
│   └── test.parquet
├── medium/
│   └── test.parquet
├── hard/
│   └── test.parquet
└── metadata/
    └── manifest.json

• easy/test.parquet: 20 easy packing cases.
• medium/test.parquet: 20 medium packing cases.
• hard/test.parquet: 20 hard packing cases.
• metadata/manifest.json: benchmark split statistics and generation metadata.

🧾 Data Fields

Each parquet row describes one packing case. The main fields include:

Field	Type	Description
data_source	string	Dataset source identifier.
prompt_template_id	string	Prompt template used by the evaluation pipeline.
difficulty	string	Difficulty level: easy, medium, or hard.
ground_truth	struct/list	Reference packing trajectory and object metadata.
buffer_size	int	Number of candidate objects visible to the policy at each step.
split	string	Dataset split, set to test.

The full schema is preserved in the released parquet files.

📊 Benchmark Size

Difficulty	Cases	Buffer Size
easy	20	3
medium	20	5
hard	20	10
Total	60	-

🎯 Intended Uses

PackLab-Bench is intended for research evaluation of robotic bin packing policies, including multimodal large language models, packing heuristics, and learning-based policies.

This benchmark is intended for evaluation rather than training.

📏 Evaluation Protocol

Each method is evaluated by executing its predicted action sequence in the PackLab environment. The benchmark reports three primary metrics:

• Success Ratio: volume ratio of successfully packed objects whose centers remain inside the container.
• Compactness: packed object volume divided by the occupied packing-envelope volume.
• Overall Score: product of Success Ratio and Compactness.

For comparability with the paper, methods should be evaluated on all 60 cases and results should be reported by difficulty and on average.

💻 How to Load

You can load the benchmark parquet files with standard Python tools:

from pathlib import Path
import pandas as pd

root = Path("PackLab-Bench")

easy = pd.read_parquet(root / "easy" / "test.parquet")
medium = pd.read_parquet(root / "medium" / "test.parquet")
hard = pd.read_parquet(root / "hard" / "test.parquet")

print(len(easy), len(medium), len(hard))
print(easy.iloc[0].to_dict().keys())

⚖️ License and Ethics

PackLab-Bench is released under CC BY-NC 4.0 for research and non-commercial benchmarking use.

Users should validate generated packing actions before deploying them on real robotic hardware. Physical execution requires appropriate safety checks, collision handling, and supervision.