Deep Reinforcement Learning For Assembly Planning

This is the official code for our paper "Deep Reinforcement Learning for Real-time Assembly Planning in Robot-based Prefabricated Construction" (Under Review).

Requirements

• pygame==2.1.2
• PyOpenGL==3.1.6
• tianshou==0.4.9.post1
• wandb

Install Packages

1. Use requirements.txt to install packages:

pip install -r requirements.txt

2. Use environment.yml to install conda environment:

conda env create -f environment.yml

Training Models

Users can use --env-id and --task-id to select the correspodning environments and scenarios:

# login wandb
wandb login
# train the agent with DQN
python run_dqn_script.py --env-id 1 --task-id 1 --epoch 100 --gamma 0.9 --seed 0
# train the agent with DDQN
python run_dqn_script.py --env-id 1 --task-id 1 --epoch 100 --gamma 0.9 --use-dueling --seed 0
# train the agent with A2C
python run_a2c_script.py --env-id 1 --task-id 1 --epoch 100 --gamma 0.9 --seed=0
# train the agent with PPO
python run_ppo_script.py --env-id 1 --task-id 1 --epoch 100 --gamma 0.9 --norm-obs --seed=0

Visualizing the Trained Agent

python demo_visualisation.py --env-id 1 --test-id 1 --algo dqn --ckpt-path <your-ckpt-path> --render

Plot Training Curves

The log files used in this paper can be downloaded from [link].

# 1. enter the plot_tools/ folder
cd plot_tools
# 2. download all log files from Env1 - Scene1
python download_data.py --user-name <wandb-user-name> --group-name Env_1_Scene_1
# 3. plot the corresponding learning curve
python extract_info.py --env-id 1 --task-id 1 --plot --logdir <path-to-save-logs>