This repo contains the implementation of physics-aware assembly sequence planning (ASP) for combinatorial Lego assembly.
This work focuses on planning physcially feasible assembly sequence for combinatorial assembly. Given the 3D shape, the policy outputs a sequence of actions for placing unit primitives to build the goal object. Importantly, the action for each step is physically feasible. Examples of planned assembly sequences are shown below.
- Configurate the config file
./config.json.Train:0: build a model.1: train the policy.data_folder: path to the dataset folder.workspace_dimension: the XYZ dimension of the building workspace.output_dir: If training, this is the directory to save the training results. If building, this is the directory to load the policy model.trial: an integer trial number.build_file_idx: an index indicating which structure to build in the dataset. No effect when training.max_step: task horizon of the ASP. No effect when building.
python3 main.py.
If you find this repository helpful, please kindly cite our work.
@article{liu2024physics,
title={Physics-Aware Combinatorial Assembly Planning using Deep Reinforcement Learning},
author={Liu, Ruixuan and Chen, Alan and Zhao, Weiye and Liu, Changliu},
journal={arXiv preprint arXiv:2408.10162},
year={2024}
}
This project is licensed under the MIT License.