Irregular-Object-Packing

Planning Irregular Object Packing via Hierarchical Reinforcement Learning.

This repo contains the basic code for Irregular Object Packing on PyBullet.

The code is based on PyTorch. Please prepare the object files from YCB and OCRTOC and put them into objects folder. In this repo there are still some objects in the folder. Use them as guideline for format requirements.

Link for downloading additional objects:

• YCB:
  • https://github.com/eleramp/pybullet-object-models
• OCRTOC:
  • http://www.ocrtoc.org

Repo organization

The repo is organized as follows:

• main_module.py: The main file script.
• env.py: Functions to interact with the PyBullet simulation environment.
• models.py: Models for sequence planning and placement planning.
• trainer.py: Trainer class to train models.
• tester.py: Tester class to test models.
• heuristics_HM.py: HM heuristic class for placement.
• environment_GPU.yml: yml file for conda environment for GPU support (NVIDIA-SMI 535.161.08 - Driver Version: 535.161.08 - CUDA Version: 12.2).
• environment_noGPU.yml: yml file for conda environment for CPU support.
• fonts_terminal.py: Colors definition for terminal prints.
• pybullet_test.py: File to test pybullet installation.
• visualize_models.py: File to visualize the models structure using graphviz.

Useful conda commands

• Create conda env from yml file

conda env create -f environment_GPU.yml

• See available environments in conda

conda env list

• export and environment in yml file

conda env export > environment.yml

Install torch

Guidelines to install torch in the conda environment according to your device setup: https://pytorch.org/get-started/locally/

Quick Start

1. The first thing to do is to create a conda environment containig the packages listed in environment_GPU.yml and environment_noGPU.yml respectively if you want to use your GPU to train/test the models or not. You can use the Useful conda commands to do that.

2. You can use the file main_module.py to train or test the models.

Train options:

• --obj_folder_path: path to the folder containing the objects .csv file
• --gui: GUI for PyBullet
• --force_cpu: Use CPU instead of GPU
• --stage: stage 1 or 2 for training
• --k_max: Max number of objects to load
• --k_min: Min number of objects to load
• --k_sort: Number of objects to consider for sorting
• --manager_snapshot: path to the manager network snapshot
• --worker_snapshot: path to the worker network snapshot
• --new_episodes: number of episodes
• --load_snapshot_manager: Load snapshot of the manager network
• --load_snapshot_worker: Load snapshot of the worker network