Install Isaac Gym on Ubuntu 22.04

Isaac Gym is a high-performance robotics simulation platform by NVIDIA, designed for creating and training intelligent robots using advanced physics simulations and deep learning. Below is a simple guide how to get started with Isaac Gym.

Install Miniconda & download Isaac Gym

mkdir -p ~/miniconda3
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda3/miniconda.sh
bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
rm -rf ~/miniconda3/miniconda.sh
~/miniconda3/bin/conda init bash

Download Isaac Gym

unzip IsaacGym_Preview_4_Package.tar.gz

tar -xf IsaacGym_Preview_4_Package.tar.gz

Install Isaac Gym

cd isaacgym
./create_conda_env_rlgpu.sh

Once it finishes, it will show SUCCESS as below:

To activate the virtual environment:

conda activate rlgpu

Test installation

cd isaacgym/python/examples
python joint_monkey.py 

Next, we will install the IsaacGymEnvs

git clone https://github.com/NVIDIA-Omniverse/IsaacGymEnvs.git
cd IsaacGymEnvs
pip install -e .

Now, let’s do some training

cd IsaacGymEnvs/isaacgymenvs
python3 train.py task=Cartpole

The program will terminate when reach max epoch

We can also set number of environment with the num_envs argument

python train.py task=BallBalance num_envs=512 train.params.config.minibatch_size=512

Troubleshooting

We may face issue like:

ImportError: libpython3.7m.so.1.0: cannot open shared object file: No such file or directory

Suggestion from the Isaac Gym documentation:

So for me, below is the solution:

export LD_LIBRARY_PATH=/home/will/miniconda3/envs/rlgpu/lib

One issue I encounter when importing torch

Downgrading MKL to 2024.0.0 resolved this issue

pip install mkl==2024.0.0

Reference:

ImportError `undefined symbol: iJIT_NotifyEvent` encountered when MKL 2024.1 is installed. · Issue…

We may also encounter

undefined symbol: cublasLtHSHMatmulAlgoInit, version libcublasLt.so.11

This happens because pytorch1.13 automatically installed nvidia_cublas_cu11, nvidia_cuda_nvrtc_cu11, nvidia_cuda_runtime_cu11 and nvidia_cudnn_cu11.

To workaround this, I did

pip uninstall nvidia_cublas_cu11

Reference:

https://stackoverflow.com/a/75095447