Closed-Loop Control of Direct Ink Writing via Reinforcement Learning

This instalation guide was tested on Ubuntu 18.04 LTS.

Custom build of GCC to remove PIE error

The code was tested with GCC 7.5.0

Prepare

sudo apt update
sudo apt upgrade
sudo apt install build-essential wget m4 flex bison

Get gcc source code wget https://ftpmirror.gnu.org/gcc/gcc-7.5.0/gcc-7.5.0.tar.xz

Unpack the source tar xf gcc-7.5.0.tar.xz

Download the prerequisities

cd ./gcc-7.5.0
./contrib/download_prerequisites

Configure languages ./configure -v --build=x86_64-linux-gnu --host=x86_64-linux-gnu --target=x86_64-linux-gnu --prefix=/usr/local/gcc-7.5.0 --enable-languages=c,c++ --disable-multilib

Build the source sudo make -j 12

Install sudo make install-strip

Change default gcc on server

sudo update-alternatives --remove-all gcc 
sudo update-alternatives --remove-all g++

sudo update-alternatives --install /usr/bin/gcc gcc /usr/local/gcc-7.5.0/bin/gcc 100
sudo update-alternatives --install /usr/bin/g++ g++ /usr/local/gcc-7.5.0/bin/g++ 100

sudo update-alternatives --config gcc
sudo update-alternatives --config g++

Install NVidia driver

Install latest nvidia driver

apt search nvidia-driver
sudo apt install nvidia-driver-455
sudo reboot

Install CUDA 9.2 on Ubuntu 18.04

Make sure you have installed the cuda driver (any version) in your system.

Install cuda dependencies sudo apt-get install freeglut3 freeglut3-dev libxi-dev libxmu-dev

Download the cuda runfile installer from: https://developer.nvidia.com/cuda-92-download-archive, select Linux > x86_64 > Ubuntu > 17.10 > runfile (local) and select Base Installer (1.7GB).

Run the installation. There are several yes/no questions, make sure to select not installing driver and install the cuda toolkit.

chmod +x cuda_9.2.148_396.37_linux.run
sudo ./cuda_9.2.148_396.37_linux.run

After installation, add these two lines into your ~/.bashrc or ~/.zshrc depending on the shell system you are using.

export PATH="$PATH:/usr/local/cuda-9.2/bin"
export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:usr/local/cuda-9.2/lib64"

Change the makefile by replacing all /usr/local/cuda/ by /usr/local/cuda-9.2/ in demo/compiler/makelinux64/Makefile.flexDemoCUDA.mk and demo/compiler/makelinux64/Makefile.flexExtCUDA.mk1. Alternatively, you can change the default cuda in the system into cuda-9.2 by

ln -s /usr/local/cuda-9.2 /usr/local/cuda

Then test the FleX

cd ./demo/compiler/makelinux64
make clean
make -j
cd ../../../bin/linux64/
./NvFlexDemoReleaseCUDA_x64

Instal EGL

apt-cache search opengles 
apt-cache search mesa | grep -i egl
sudo apt-get install mesa-utils-extra
sudo apt-get install libegl1-mesa
sudo apt-get install libegl1-mesa-dev

Build libpyflex

cd ./libpyflex
./build.sh

The build creates a dynamic library pyflexlib.so that should work on both your local machine and a server node.

Setting up conda environment

Install conda

Set the path in environment.yml by replacing the <username> with your user name.

Load the predefined conda environment conda env create -f environment.yml

Verify if rl_printerenvironment is present conda list

Set Up Learning Module

cd learning/externals/
cd baselines
pip install -e .
cd ../pytorch-a2c-ppo-acktr-gail
pip install -e .

Example scenes

./learning/train/test_pyflex.py

Initializes 5 Flex environments and starts the deposition in place 3 times. The tests confirms that the library was build succesfully and that the simulation works. Upon execution this image should be created:

./learning/train/test_printer.py

Starts a printing process controlled through python. The printer follows a lucky leaf print. After printing the following image should be created:

Start a Learning Process

cd ./learning/train
python cnn_reinforce.py

Starts the learning process for outline printing. The training process will be stored in trained_models folder under a unique id. Each process contains:

• args.txt a list of arguments for the training
• logs.txt a log of reward and other useful debug info
• models/ a folder with all intermediate models discovered during the learning process

Visualize the Learning

python training_visualize.py --log-path=./trained_models/FlexPrinterEnv-v0/<unique_id>/logs.txt

Displays the per-iteration reward logged during training.

Visualize Trained Controllers

python collect_training.py --model-path=./trained_models/FlexPrinterEnv-v0/<unique_id>/models/FlexPrinterEnv-v0_iter

Runs the controller for the entire slice dataset. To select the controller set the id in collect_training.py .

Pyflexlib API

pyflex.init(width, height, renderDevice = 0, border = 9.0)

Creates a new pyflex instance. The instances will render into an image of pixel size width x height using the selected rendering device (for multi-gpu systems). The border sets the physical dymensions of the image in milimeters.

pyflex.delete(width, height, renderDevice = 0, border = 9.0)

Releases the pyflex instance from the memory.

pyflex.reset(numSubsteps, materialViscosity, materialReservoir)

Resets the simulation and erases all the previously deposited material. Sets the material properties and the amount of available material.

pyflex.step(x, y, z, flow)

Performs one simulation step by translating the printing nozzle to location [x,y,z] while depositing materials with a given flow.

Printer Environment API

The printer environment implements the interface of OpenAI gym and is fully compatible.