conda create -n prompt python==3.6.8
conda activate prompt
pip install -r requirements.txt
conda install pybind11
conda install pytorch torchvision cudatoolkit=11.1 -c pytorch -c nvidia

pip install -e gym_env/

docker pull adacomplab/prompt_flex
docker run \
  -v "$(pwd)"/PyFlexEngine/:/workspace/PyFleX \
  -v ~/anaconda3/envs/prompt/:/workspace/anaconda \
  -it adacomplab/prompt_flex
  
# in the docker image, to build prompt_pyflex
cd ~
export PATH="/workspace/anaconda/bin:$PATH"
cd /workspace/PyFleX
export PYFLEXROOT=${PWD}
export PYTHONPATH=${PYFLEXROOT}/bindings/build:$PYTHONPATH
export LD_LIBRARY_PATH=${PYFLEXROOT}/external/SDL2-2.0.4/lib/x64:$LD_LIBRARY_PATH
cd bindings; mkdir build; cd build; cmake ..; make -j

export PYFLEXROOT=Path_To_Prompt/PyFlexEngine
export PYTHONPATH=${PYFLEXROOT}/bindings/build:$PYTHONPATH
export LD_LIBRARY_PATH=${PYFLEXROOT}/external/SDL2-2.0.4/lib/x64:$LD_LIBRARY_PATH

# start planning service
# you may reduce the num_workers, if out of memory occurs
# you can set "--visualize_flex 0" to run planning headlessly and speed up planning.
python core/planning/grasp/grasp_plan_service.py --gpu_id 0 --num_workers 20 --visualize_flex 1
# if you have multiple gpus
# CUDA_VISIBLE_DEVICES=1 python core/planning/grasp/grasp_plan_service.py --gpu_id 1 --num_workers 20
# CUDA_VISIBLE_DEVICES=2 python core/planning/grasp/grasp_plan_service.py --gpu_id 2 --num_workers 20
# start grasping demo
# you can set "--visualize_pybullet 0" to run robot simulation headlessly in pybullet.
# grasping screenshots are also saved in core/app/grasp/data_folder/
python core/app/grasp/main_grasp.py --num_gpu 1 --visualize_pybullet 1