PackLab-VLM-9B

📌 Checkpoint Summary

PackLab-VLM-9B is a packing-specialized vision-language model for closed-loop robotic bin packing. It is fine-tuned from a Qwen3.5-9B backbone with packing-oriented supervised fine-tuning on PackData-20K.

The model predicts structured packing actions from multimodal packing states, including the container heightmap, candidate-object attributes, and observation-action history.

💻 How to Use

Please refer to the PackLab code repository for environment setup, inference, evaluation, and data formatting instructions.

By default, the released code expects the checkpoint at:

weights/PackLab-VLM-9B/

The checkpoint is released in HuggingFace format and contains:

PackLab-VLM-9B/
├── README.md
├── chat_template.jinja
├── config.json
├── generation_config.json
├── model.safetensors
├── processor_config.json
├── tokenizer.json
└── tokenizer_config.json

🎯 Intended Uses

This checkpoint is intended for research on multimodal robotic bin packing, closed-loop packing policies, and physical packing evaluation. Users should evaluate generated actions carefully before applying the model to physical robotic systems.

⚖️ License and Ethics

The checkpoint is released under the Apache License 2.0.

Users should apply the model responsibly and avoid unsafe robotic deployment without appropriate validation, supervision, and hardware safeguards. Packing actions are determined by the model input, model weights, and runtime decoding settings.