AirGym (Still Developing)

This project is based on Aerial Gym Simulator licensed(https://github.com/ntnu-arl/aerial_gym_simulator) under the BSD 3-Clause License. It has been modified and greatly improved by emNavi Technology.

This project provides a realistic dynamics and RL framework for Sim2Real tasks of quadcopter. Drones can be trained in AirGym and then transferred to reality.

We also build a randomly generated environment for drone training. One demo is shown below:

Requirements

1. Ubuntu 20.04 or 22.04
2. Conda or Miniconda
3. NVIDIA Isaac Gym Preview 4 (Pytorch needs to upgrade for 40 series of GPUs. Please follow the installation guidance.)

Note this repository has been tested on Ubuntu 20.04/22.04 with PyTorch 2.0.0 + CUDA11.8.

Installation

1. Install IsaacGym Preview 4 with torch2.0.0+cuda11.8

1. Download package from the official page and unzip.
2. Edit install_requires in python/setup.py:

   install_requires=[
               "numpy",
               "scipy",
               "pyyaml",
               "pillow",
               "imageio",
               "ninja",
           ],
   

3. Edit dependencies in python/rlgpu_conda_env.yml:

   dependencies:
   - python=3.8
   - numpy=1.20
   - pyyaml
   - scipy
   - tensorboard

4. Create a new conda environment named rlgpu and install isaacgym:

   cd isaacgym
   ./create_conda_env_rlgpu.sh

5. Install PyTorch2.0.0 and CUDA11.8:

   conda activate rlgpu
   conda install pytorch==2.0.0 torchvision==0.15.0 torchaudio==2.0.0 pytorch-cuda=11.8 -c pytorch -c nvidia

2. Install rlPx4Controller

1. Install Eigen (The recommend version is 3.3.7):

   sudo apt install libeigen3-dev

2. Install pybind11:

   pip install pybind11

3. Install rlPx4Controller

   git clone git@github.com:emNavi/rlPx4Controller.git
   cd rlPx4Controller
   pip install -e .

3. Install AirGym

git clone git@github.com:FP-Flight/AirGym.git
cd AirGym/
pip install -e .

4. Test the installation

Run the example script:

cd airgym/scripts
python example.py --controller_test --task X152b --ctl_mode pos

The default ctl_mode is position control.

Training and Displaying

We can train the model by rl-games==1.6.1 which is a dependency of airgym. Of course you can use other RL libs for training. emNavi provide a set of quadrotor sim2real approach by using rl_games, please refer to AirGym-Real @https://github.com/emNavi/AirGym-Real.

Training:

cd airgym/rl_games/
python runner.py --ctl_mode rate --headless

Algorithm related parameters can be edited in .yaml files. Environment and simulator related parameters are located in ENV_config files like X152bPx4_config.py. The ctl_mode must be spicified.

Displaying:

cd airgym/rl_games/
python runner.py --play --num_envs 64 --ctl_mode rate --checkpoint <path-to-ckpt>

Training a Trajectory Tracking Policy

Every task is mainly affected by two .py files. Use task X152b_sigmoid as an example. Env definition file is X152b_sigmoid.py, and the config file is X152b_sigmoid_config.py, which could change environmental configuration like control mode, adding assets, simulation specification. ctl_mode has five options: 'pos', 'vel', 'atti', 'rate', 'prop', and 'pos' is the default setting.

Algorithm related configuration can be edited in ppo_X152b_sigmoid.yaml. Environment related configuration can be edited in .../envs/.../X152b_sigmoid_config.py.

Training:

cd airgym/rl_games/
python runner.py --task X152b_sigmoid --headless --ctl_mode rate  --file ppo_X152b_sigmoid.yaml

Displaying:

cd airgym/rl_games/
python runner.py --play --num_envs 4 --task X152b_sigmoid --ctl_mode rate --checkpoint <path-to-model>