Skip to content

Latest commit

 

History

History
108 lines (83 loc) · 3.42 KB

INSTALL.md

File metadata and controls

108 lines (83 loc) · 3.42 KB

Installation Instructions

  1. Install general python packages
conda create -n gembench python==3.10

conda activate gembench

# On CLEPS, first run `module load gnu12/12.2.0`

conda install nvidia/label/cuda-12.1.0::cuda
pip install torch==2.3.0 torchvision==0.18.0 torchaudio==2.3.0 --index-url https://download.pytorch.org/whl/cu121

export CUDA_HOME=$HOME/miniconda3/envs/gembench
export CPATH=$CUDA_HOME/targets/x86_64-linux/include:$CPATH
export LD_LIBRARY_PATH=$CUDA_HOME/targets/x86_64-linux/lib:$LD_LIBRARY_PATH
export PATH=$CUDA_HOME/bin:$PATH

pip install -r requirements.txt

# install genrobo3d
pip install -e .
  1. Install RLBench
mkdir dependencies
cd dependencies

Download CoppeliaSim (see instructions here)

# change the version if necessary
wget https://www.coppeliarobotics.com/files/V4_1_0/CoppeliaSim_Edu_V4_1_0_Ubuntu20_04.tar.xz
tar -xvf CoppeliaSim_Edu_V4_1_0_Ubuntu20_04.tar.xz

Add the following to your ~/.bashrc file:

export COPPELIASIM_ROOT=$(pwd)/CoppeliaSim_Edu_V4_1_0_Ubuntu20_04
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$COPPELIASIM_ROOT
export QT_QPA_PLATFORM_PLUGIN_PATH=$COPPELIASIM_ROOT

Install Pyrep and RLBench

git clone https://github.com/cshizhe/PyRep.git
cd PyRep
pip install -r requirements.txt
pip install .
cd ..

# Our modified version of RLBench to support new tasks in GemBench
git clone https://github.com/rjgpinel/RLBench
cd RLBench
pip install -r requirements.txt
pip install .
cd ../..
  1. Install model dependencies
cd dependencies

# Please ensure to set CUDA_HOME beforehand as specified in the export const of the section 1
git clone https://github.com/cshizhe/chamferdist.git
cd chamferdist
python setup.py install
cd ..

git clone https://github.com/cshizhe/Pointnet2_PyTorch.git
cd Pointnet2_PyTorch/pointnet2_ops_lib
python setup.py install
cd ../..

# llama3: needed for 3D-LOTUS++
git clone https://github.com/cshizhe/llama3.git
cd llama3
pip install -e .
cd ../..
  1. Running headless

If you have sudo priviledge on the headless machine, you could follow this instruction to run RLBench.

Otherwise, you can use singularity or docker to run RLBench in headless machines without sudo privilege. The XVFB should be installed in the virtual image in order to have a virtual screen.

A pre-built singularity image can be downloaded here. Here are some simple commands to run singularity:

export SINGULARITY_IMAGE_PATH=`YOUR PATH TO THE SINGULARITY IMAGE`
export python_bin=$HOME/miniconda3/envs/gembench/bin/python

# interactive mode
singularity shell --bind $HOME:$HOME,$SCRATCH:$SCRATCH --nv $SINGULARITY_IMAGE_PATH

# run script
singularity exec --bind $HOME:$HOME,$SCRATCH:$SCRATCH --nv $SINGULARITY_IMAGE_PATH xvfb-run -a ${python_bin} ...
  1. Adapt the codebase to your environment

To adapt the codebase to your environment, you may need to modify the following:

  • replace everywhere $HOME/codes/robot-3dlotus with your path to robot-3dlotus folder
  • replace everywhere the sif_image path to your path to the singularity image nvcuda_v2.sif