gym-xarm6 is a gym environment for path planning with xArm6 robot using openai gym framework. The environment Creation of a new environment in OpenAI gym for the xArm6 robot from UFactory. The model uses Deep Deterministic Policy Gradient (DDPG) for continious actions and Hindsight Experience Replay (HER).

• python3.6+ environment by one of the following:
  • conda
  • virtualenv
  • venv
  • system python
• swig
• libosmesa6-dev
• patchelf
• libopenmpi-dev
• mpi4py (install it with pip or conda)
• openai baselines

Follow this instructions to be able to use the xArm6 enviroment for OpenAI gym. You can use your own policies to train the model, but here we will use DDPG + HER.

Before the installation, we need to get MuJoCo (multi-joint dynamics in contact) physics simulator.

Note: For easier installation we recomend using Unix based OS.

1. The code was tested on Ubuntu 20.04, Debian 11 succesfully.
2. Have fun manipulating the code and the model!

1. Download MuJoCo
   • Windows
   • Linux
   • MacOS
2. Create a folder in your home directory named .mujoco and unzip the file here.
3. Download the activation key licence and put it inside .mujoco/mujoco200/ and .mujoco/mujoco200/bin/.
4. Modify the .bashrc and add the following lines:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/your/home/dir/.mujoco/mujoco200/bin

5. Test that MuJoCo is working properly.

cd ~/.mujoco/mujoco200/bin
./simulate ../models/humanoid.xml

Now clone this repository.

git clone https://github.com/ollintzinlab/xArm6-Gym-Env.git

Once you have it, cd into the cloned directory an run pip install -e . to install the environment.

Train the model by running:

python her/train.py --env-name="xArm6Reach-v0" | tee her/reach.log

Test the model by running:

python her/demo.py --env-name="xArm6Reach-v0"

In gym_xarm6/envs/reach.py you can modify the position of the robot in the world frame.

...
        initial_qpos = {
            'robot0:slide0': 0.,
            'robot0:slide1': 0.,
            'robot0:slide2': 0.,
        }
...

The variable distance_threshold allows you to change de precision of the robot. It is used to calculate the normal distance from the tool frame of the robot to the target position in a cartesian frame.