donghao-zhou/PackData-20K

📌 Dataset Summary

PackData-20K is the supervised fine-tuning dataset used to train PackLab-VLM for closed-loop robotic bin packing.

Each sample contains a multimodal packing state and the corresponding structured action target. The model observes the current container heightmap, candidate-object information, and observation-action history, then learns to select an object, orientation, and target placement position.

🗂️ File Structure

The released dataset contains the training split only:

PackData-20K/
├── README.md
├── train/
│   ├── easy/
│   │   ├── shard_00000_of_00100/train.parquet
│   │   └── ...
│   ├── medium/
│   │   ├── shard_00000_of_00100/train.parquet
│   │   └── ...
│   └── hard/
│       ├── shard_00000_of_00100/train.parquet
│       └── ...
├── images/
│   ├── easy/
│   │   ├── shard_00000_of_00100.tar
│   │   └── ...
│   ├── medium/
│   │   ├── shard_00000_of_00100.tar
│   │   └── ...
│   └── hard/
│       ├── shard_00000_of_00100.tar
│       └── ...
└── metadata/
    └── metadata_summary.json

The validation split used during internal training is not included in this release.

🧾 Data Fields

Each parquet row contains one supervised training example:

Field	Type	Description
data_source	string	Dataset source identifier.
prompt_template_id	string	Prompt template used to format the training sample.
enable_thinking	bool	Whether thinking-style generation is enabled for the sample.
messages	list	Chat-format user and assistant messages.
images	list	Relative paths to heightmap images used by the multimodal prompt.
extra_info	struct	Packing metadata, including difficulty, buffer size, object dimensions, target plan, and trajectory statistics.

Image paths inside the parquet files are relative to the dataset root. After extracting the image archives, paths such as the following should resolve directly:

images/train/easy/shard_00000_of_00100/easy_000000_step_000.png

📊 Data Size

Split	Difficulty	Samples
train	easy	9,000
train	medium	7,000
train	hard	4,000
train	total	20,000

The training split is stored in 300 parquet shards, with 100 shards per difficulty.

🖼️ Image Archives

The heightmap images are packaged as shard-level tar archives to avoid uploading hundreds of thousands of small files:

images/easy/shard_00000_of_00100.tar
images/medium/shard_00000_of_00100.tar
images/hard/shard_00000_of_00100.tar

Extract the archives from the dataset root before training:

for difficulty in easy medium hard; do
  for archive in images/${difficulty}/*.tar; do
    tar -xf "${archive}"
  done
done

After extraction, the image directory should contain:

images/train/easy/
images/train/medium/
images/train/hard/

💻 How to Load

You can load the parquet files with standard Python tools:

from pathlib import Path
import pandas as pd

root = Path("PackData-20K")
sample_file = root / "train" / "easy" / "shard_00000_of_00100" / "train.parquet"

df = pd.read_parquet(sample_file)
sample = df.iloc[0].to_dict()

image_path = root / sample["images"][0]["image"]
print(sample["messages"])
print(image_path)

🎯 Intended Uses

PackData-20K is intended for research on supervised fine-tuning of multimodal policies for robotic bin packing. It can be used to train models that produce structured packing actions from visual and textual packing states.

⚖️ License and Ethics

PackData-20K is released under CC BY-NC 4.0 for research and non-commercial use.

Users should validate learned policies carefully before applying them to real robotic hardware. Physical deployment requires appropriate safety checks, collision handling, and supervision.