Skip to content

Latest commit

 

History

History
60 lines (47 loc) · 2.83 KB

README.md

File metadata and controls

60 lines (47 loc) · 2.83 KB
Raw

Reinforcement Learning for Sequential Assembly of SL-Blocks

This repository contains Grasshopper files and Python scripts for running sequential assembly experiments from the paper "Reinforcement Learning for Sequential Assembly of SL-Blocks". Reinforcement learning library Stable-Baselines3 is used for training the agents, while the environments are implemented in Grasshopper — an algorithmic modeling tool for Rhino.

Installation

Required Rhino/Grasshopper plugins can be downloaded from Food4Rhino and installed as described on their respective pages

  • Hoopsnake
  • Matrixtoolbox
  • Pufferfish
  • Treesloth
  • Yellow
  • Zebra

If the plugin Yellow is not available, file yellow.gha provided here in Grasshopper Files can be used as a substitute. As is common for all plugins, you may need to check the box "Unblock" in the file properties to prevent Windows from blocking it.

For a general description of the interface between Grasshopper and Stable-Baselines3, see Grasshopper Gym. It is recommended to use Anaconda to get a running Python installation under Windows.

Training and Evaluating Agents

The basic script train-and-eval.py provides the code for training and evaluating agents. To run the training, execute the command

python train-and-eval.py --algorithm="PPO" --save-path="DESIRED_PATH"

Here you need to specify the algorithm (either PPO or DQN) and the path where the model should be saved. Later, to evaluate the saved model, execute the command

python train-and-eval.py --algorithm="PPO" --load-path="DESIRED_PATH"

Before executing these commands, start the Hoopsnake loop inside the Grasshopper file Reinforcement Learning Interface.gh located in the Grasshopper Files directory. Parameters of the reinforcement learning algorithms can be adjusted directly in the train-and-eval.py script where DQN and PPO objects are created (lines 92–102).

Experiments on 2D and 3D Curves

Python scripts for training DQN and PPO agents on the 2D and 3D curves described in the paper are provided in the folder Python Files with the parameters are already pre-specified. Detailed instructions on how to use those scripts and how to modify the Grasshopper environment in Reinforcement Learning Interface.gh can be found in the accompanying video Reinforcement Learning for Sequential Assembly of SL-Blocks in Grasshopper. Other algorithms used as baselines in the paper are implemented directly in Grasshopper in the file Evolutionary Algorithm and Greedy Algorithm for SL-Blocks.gh. Pre-trained weights for DQN and PPO agents are available in the directories DQN-Experiments and PPO-Experiments.