JPack:

JPack is a Jittor-based reconstruction project for Online 3D Bin Packing (Online 3D-BPP). Built on top of our team’s previously released RoboBPP benchmark, this repository ports several representative Torch-based methods to Jittor and organizes them into a reusable codebase for training, testing, and evaluation.

The repository currently contains four core subprojects:

• AR2L: Adjustable Robust Reinforcement Learning
• PCT: Packing Configuration Tree
• PCT-full: an extended PCT framework with additional packing models and evaluation utilities
• TAPNet++: an end-to-end framework from perception to packing decision

Benefiting from Jittor’s meta-operator system and kernel fusion capabilities, JPack provides a lightweight and efficient reconstruction for industrial online packing research. According to our project note, the reconstructed models achieve an average 10% improvement in forward inference speed, making them better suited for real-time decision-making in dynamic production environments.

Overview

Online 3D bin packing requires an algorithm to decide the placement position and pose of each incoming item without knowing the future item sequence. In real industrial scenarios, the problem is not only about maximizing space utilization, but also about balancing:

• stacking stability
• physical executability
• robotic operation safety

Many previous studies rely on idealized geometric assumptions. In contrast, RoboBPP pushes the problem toward a more realistic benchmark setting with physical constraints and robotic execution constraints. JPack further extends this effort by providing Jittor reconstructions of representative learning-based methods in the RoboBPP benchmark, supporting:

• unified algorithm comparison and reproducible experiments
• training and testing on real industrial data
• progressive evaluation from pure geometric packing to physics simulation and execution-aware testing

Benchmark Setting

Real-World Industrial Data

JPack follows RoboBPP and targets three representative industrial data scenarios:

• Repetitive Dataset: assembly-line style scenarios with highly repetitive item patterns
• Diverse Dataset: warehouse and e-commerce style scenarios with highly heterogeneous item sizes
• Wood Board Dataset: long-board and panel scenarios that require stronger stability and pose control

In the current codebase and example commands, common dataset paths include:

• data/time_series/pg.xlsx
• data/occupancy/deli.xlsx
• data/flat_long/opai.txt

Progressive Evaluation

JPack inherits the staged evaluation design of RoboBPP and covers three levels of testing:

• Math Pack: upper-bound performance under ideal geometric assumptions
• Physics Pack: stability evaluation with gravity, friction, and other physical constraints
• Execution Pack: evaluation with additional robotic grasping and execution constraints

This evaluation protocol is closer to real industrial deployment and highlights the gap between geometric feasibility and physical executability.

Repository Structure

JPack/
├── AR2L/        # Jittor reconstruction of adjustable robust RL
├── PCT/         # Jittor reconstruction of Packing Configuration Tree
├── PCT-full/    # Extended PCT framework with more model variants and evaluation scripts
├── TAPNet++/    # Jittor implementation of TAPNet++
├── Jpack.tex    # Project note / communication draft
└── README.md

The four subdirectories serve the following purposes:

• AR2L: robust online packing policy learning under challenging or adversarial item sequences
• PCT: hierarchical packing state representation and reinforcement learning based on Packing Configuration Trees
• PCT-full: a broader experimental framework with more model architectures and benchmark-oriented evaluation entry points
• TAPNet++: joint packing decision optimization with candidate space representation and policy learning

Installation

We recommend using Python 3.8. Since the four subprojects do not share exactly the same dependency set, it is recommended to create separate environments for different methods when necessary, or install dependencies on demand.

conda create -n jpack python=3.8
conda activate jpack
pip install jittor==1.3.10.0

Then install the dependencies required by the target subproject:

• AR2L/requirements.txt
• PCT/requirements.txt
• PCT-full/requirements.txt
• TAPNet++/requirements.txt

Notes:

• AR2L, PCT, and PCT-full mainly depend on the gym 0.x ecosystem
• TAPNet++ uses gymnasium and tianshou
• PCT-full additionally depends on physics-related packages such as pybullet and trimesh

Quick Start

Run the following commands inside the corresponding subdirectory. The example arguments are adapted from the existing demo.sh files or experiment scripts. Please replace the dataset path with the valid JPack / RoboBPP data files on your machine.

1. PCT

cd PCT
python main.py --setting 1 --internal-node-holder 80 --leaf-node-holder 50 --load-dataset --dataset-path data/time_series/pg.xlsx --custom time_series --container-size 134 125 100

2. AR2L

cd AR2L
python main.py --num-box 80 --num-next-box 20 --num-candidate-action 120 --alpha 1 --dataset-path data/time_series/pg.xlsx --custom time_series

3. PCT-full

cd PCT-full
python main.py --setting 2 --custom discrete_real_opai_setting_2 --preview 1 --select 1 --internal_node_holder 150 --leaf_node_holder 150 --env_version 3 --device 0 --item_size_set discrete --container_size 250,120,100 --training_without_evaluate --distribution real_opai

PCT-full also provides:

• batch experiment examples: run_ijrr.py
• offline evaluation entry: evaluation.py
• multiple model variants: PCT, CDRL, PackE, Attend2Pack, RCQL, and Rainbow

4. TAPNet++

cd TAPNet++
python tap_train.py --box-num 30 --box-range 10 80 --container-size 250 120 100 --train 1 --test-num 1 --model tnpp --prec-type none --fact-type box --data-type rand --ems-type ems-id-stair --stable-rule none --rotate-axes z --hidden-dim 128 --world-type ideal --container-type single --pack-type last --stable-predict 0 --note 4corner --corner-num 1 --max-epoch 300 --reward-type H --dataset-path data/flat_long/opai.txt --custom flat_long

For testing, change --train 1 to --train 0.

Included Methods

This repository reconstructs and organizes several representative methods used in the RoboBPP benchmark:

• PCT: represents packing states as Packing Configuration Trees and reduces continuous placement difficulty through candidate-node selection
• TAPNet++: jointly optimizes item order, orientation, and placement position in a unified decision step
• AR2L: explicitly models the trade-off between average-case performance and worst-case robustness
• PCT-full: extends the unified framework with more model architectures and benchmark-oriented evaluation utilities

Use Cases

JPack is suitable for:

• reproducing Online 3D Bin Packing reinforcement learning methods with Jittor
• benchmarking algorithms under a RoboBPP-style evaluation protocol
• studying the transition from ideal geometric packing to physics-based and execution-aware evaluation
• embodied intelligence research for industrial logistics, warehousing, and board-manufacturing scenarios

Citation

If you use JPack in your research, please cite both the RoboBPP benchmark and the corresponding algorithm papers:

@inproceedings{zhao2021learning,
  title={Learning Efficient Online 3D Bin Packing on Packing Configuration Trees},
  author={Zhao, Hang and Yu, Yang and Xu, Kai},
  booktitle={International Conference on Learning Representations},
  year={2021},
}

@article{xu2023neural,
  title={Neural Packing: From Visual Sensing to Reinforcement Learning},
  author={Xu, Juzhan and Gong, Minglun and Zhang, Hao and Huang, Hui and Hu, Ruizhen},
  journal={ACM Transactions on Graphics},
  volume={42},
  number={6},
  pages={1--11},
  year={2023},
  publisher={ACM New York, NY, USA}
}

@article{pan2023adjustable,
  title={Adjustable robust reinforcement learning for online 3d bin packing},
  author={Pan, Yuxin and Chen, Yize and Lin, Fangzhen},
  journal={Advances in Neural Information Processing Systems},
  volume={36},
  pages={51926--51954},
  year={2023}
}

@article{wang2025robobpp,
  title={RoboBPP: Benchmarking Robotic Online Bin Packing with Physics-based Simulation},
  author={Wang, Zhoufeng and Zhao, Hang and Xu, Juzhan and Zhang, Shishun and Xiong, Zeyu and Hu, Ruizhen and Zhu, Chenyang and Zeng, Zecui and Xu, Kai},
  journal={arXiv preprint arXiv:2512.04415},
  year={2025}
}

Acknowledgements

This project is built on top of our team’s previously proposed RoboBPP benchmark and reconstructs the core implementations of several representative online packing methods. It is contributed by members of a joint team from:

• Institute of AI For Industries, Chinese Academy of Sciences
• Shenzhen University
• National University of Defense Technology
• Wuhan University
• China Post Technology
• JD Technology
• SF Technology

We would also like to thank:

• the authors of RoboBPP, PCT, AR2L, TAPNet++, and related works
• the Jittor framework for supporting efficient reconstruction and deployment