3D-Bin-Packing Project

This project is a 3D Packing Optimization project under mentorship of InstaDeep. The source code of this repository can be found at Hang Zhao. The ultimate goal is to pack a Unit Load Device Unit Load Device Description

image credit Marc Lacoste

Specific dimensions for several Unit Load Devices.

Why?

Is it possible to pack shipping containers more efficiently? Save space, save time, save money - pack efficiently. Packing efficiently has the potential to reduce shipping costs that have increased due to the increase in fuel costs, shipping containers, trucks, and shipping bottlenecks.

Scope

Open-ended research questions:

• Which package do I pick next and how do I set the order of items before placing/packing?
• For an item (id: 1, length: L, width: W, height: H): Where do I place it and do I rotate it? Do I force the first package to be in the bottom-left corner?
• How do I observe and encode the current state?

Expected Learning Outcomes

1. Building a clean RL environment in python
2. Code testing and possibly test-driven development (TDD)
3. Learning to formulate real-world use-cases into ML/RL problems
4. Learning to implement and/or use advanced ML/RL algorithms and models

How to get it work

1. Prepare environment

git clone [email protected]:sangnguyens/3D-Bin-Packing.git
cd 3D-Bin-Packing
conda env create --file packing_env.yml

2. Train model (default mode):

python main.py

3. Evaluate model from pretrained model (default mode):

python evaluation.py

Rendering

Run evaluation.py and come to nb folder run BinRenderingCode.ipynb to see how agent packs and places boxes.

Help

python main.py -h
python evaluation.py -h

Our Experiments

For our experiments we trained the RL agents for 35,000 updates and updated and saved the model every 1,000 updates. We used the following parameters to train our agent:

• Trained for a total of 35,000 updates.
• Updated and saved the model every 1,000 updates using flags --model-save-interval 1000 --model-update-interval 1000.
• Generated data that sampled boxs with edge lengths varying from 1 to 5 using the file givenData.py by editing the lower and higher edges of item_size_set.
• Trained using a bin size of either 10x10x10 or 20x10x10 by editing the container_size in givenData.py.
• A random seed of 42 with the flag --seed 42.
• Set the internal node holder to 800 to prevent the model from exiting training early with the flag --set internal-node-holder 800.
• Used an either A2C or acktr reinforcement learning agent using the flag -use-acktr USE_ACKTR to select the acktr agent as the A2C agent is used by default.

To reproduce, the general run command was similar to:

python main.py --seed 42 --internal-node-holder 800 --use-acktr USE_ACKTR --model-update-interval 1000  --model-save-interval 1000

We then evaluated the different trained models using several different selections of boxes by editing givenData.py:

• Uniform 2x2x2 cubes.
• The same generated data sampling.
• Long skinny boxes by sampling from boxes generated with 2 short edges (less than 3) and one long edge (greater than 5)
• Large flat boxes by sampling from boxes generate with 1 short edge (less than 3) and two long edges (greater than 5)

To reproduce, the general evaluation command was similar to:

python evaluation.py --evaluate --internal-node-holder 800  --load-model --model-path 'Path-to-saved-model/model.pt

Group Members

• Jongbum Lee
• Cristina Moody
• Sang Nguyen

Crediting

Hang Zhao paper

@inproceedings{
zhao2022learning,
title={Learning Efficient Online 3D Bin Packing on Packing Configuration Trees},
author={Hang Zhao and Yang Yu and Kai Xu},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=bfuGjlCwAq}
}