We develop a learning-based solver for packing arbitrarily-shaped (irregular) objects in a physically realizable problem setting, which is arguably the most challenging setting of bin packing problems. This work is newly been accepted by ACM Transactions on Graphics (TOG). See these links for video demonstration: YouTube, bilibili
We release our source code and build well-established benchmark datasets. Our datasets consist of training and testing objects, separated into regular cube sets and large-scale irregular sets with various geometric characteristics, and specifying a container size.
As our reviewers suggested, although there have been numerous packing papers using RL for higher packing density, there is a lack of a common dataset to benchmark performance. We believe that having such a common benchmark would greatly facilitate research and comparison of different techniques.
If you are interested, please star this repo!
For more details, please see our paper Learning Physically Realizable Skills for Online Packing of General 3D Shapes. If this code is useful for your work, please cite our paper:
@article{zhao2022learning,
title={Learning Physically Realizable Skills for Online Packing of General 3D Shapes},
author={Zhao, Hang and Pan, Zherong and Yu, Yang and Xu, Kai},
journal={ACM Transactions on Graphics},
year={2022},
publisher={ACM New York, NY}
}
We provide the packing utility performance of our method on each dataset here for quick comparison. We run all methods in the same environment setup with the same test sequences. A total of 2000 object sequences, randomly generated from each dataset, are tested.
| Dataset | Online | Buffered (k = 3) | Buffered (k = 5) | Buffered (k = 10) |
|---|---|---|---|---|
| General | 44.5% | 45.3% | 47.5% | 55.7% |
| BlockOut | 71.0% | 71.7% | 74.8% | 78.0% |
| Kitchen | 39.2% | 44.3% | 48.8% | 57.8% |
| ABC | 34.0% | 35.3% | 37.4% | 41.0% |
| Cube | 66.1% | 67.9% | 70.8% | 74.8% |
Note: The cube dataset is composed of cubes with dimensions ranging from 0.03 to 0.15, with a unit size of 0.03. During training, we set action resolution (resolutionA) and heightmap resolution (resolutionH) to be 0.01.
- Please install dependencies by running this:
conda create -n irbpp python==3.7.7
conda activate irbpp
pip install -r requirements.txt- This code is verified on OS: Ubuntu 16.04.
For training online IR-BPP on the blockout dataset (mentioned in our paper) with our method and the default arguments:
python main.py The training data is generated on the fly. The training logs (tensorboard) are saved in './logs/runs'. Related file backups are saved in './logs/experiment'.
If you need to adjust the container size or the height map resolution, so that the size of discretized height map changes, you should adjust the CNN in 'model.py' so that it can take the modified height map as input.
You can download the prepared dataset from here. Unzip the downloaded datasets and place them into 'IR-BPP/dataset'. Each dataset includes:
shape_vhacd (folder): Watertight convex decompositions of each polygonal mesh shape. Used for accurate physics simulation.
pointCloud (folder): Point clouds pre-sampled from mesh surface. The DRL agent will re-sample 1024 points and pass them to a lightweight pointNet for shape representation.
test_sequence.pt: The dataset consists of 10000 randomly generated trajectories, each with 100 shapes. Each item with a specified planar-stable pose is presented as an id, recorded in 'id2shape.pt'.
id2shape.pt: Map the shape id to the shape name.
We provide specific tools for generating the dataset compatible with our code. If you need the corresponding undecomposed mesh files, for rendering (like our teaser) and paper writing needs, download from here.
Potential collaborations or any questions, please feel free to contact me (alex.hang.zhao@gmail.com).
We provide pre-trained models here.
For training online 3D BPP instances on BlockOut:
python main.py --device 0 --dataset blockout --custom blockoutexp --bufferSize 1 --num_processes 16 --distributed --samplePointsNum 1024 --selectedAction 500 --resolutionA 0.02 --resolutionH 0.01If you want to train a model that works on the buffered case with buffer size 10, add '--hierachical' and '--bufferSize 10'':
python main.py --device 1 --dataset blockout --custom blockoutexphier --hierachical --bufferSize 10 --num_processes 16 --distributed --samplePointsNum 1024 --selectedAction 500 --resolutionA 0.02 --resolutionH 0.01To evaluate an online model, you can add the --evaluate flag to main.py, and specify your preferred location model in --locmodel, for example:
python main.py --evaluate --evaluation-episodes-test 2000 --locmodel path/to/your/location/model --device 0 --data_name blockout --custom blockoutexp --bufferSize 1 --num_processes 1 --samplePointsNum 1024 --selectedAction 500 --resolutionA 0.02 --resolutionH 0.01If you want to test models with buffer, you should add --hierachical and also specify your preferred order model in --ordmodel, for example:
python main.py --evaluate --evaluation-episodes-test 2000 --hierachical --locmodel path/to/your/location/model --ordmodel path/to/your/order/model --data_name blockout --custom blockoutexp --bufferSize 10 --num_processes 1 --samplePointsNum 1024 --selectedAction 500 --resolutionA 0.02 --resolutionH 0.01python main.py -hThis source code is released only for academic use. Please do not use it for commercial purposes without authorization of the author.