Integrate NICOL in OpenAI python env

Adapted gazebo_nicol_v1.py gym env from ideas_deep_rl repo. Integrated via nicol python interface.

ROS melodic and a dedicated catkin workspace, e.g. ~/catkin_ws are needed. You can follow this setup guide, it also contains links to the additional ROBOTIS open_manipulator packages which you will need to start NICOL. However, requirements vary and the ReadMe is still under development, if you have any trouble installing the repo just write a comment or contact me by e-mail. NOTE: You need to check out nias (link to branch) and nias-api (link to branch) repo. Please use the linked branches.

Currently a reacher task is implemented, where NICOL must reach for a point on the table-top of it's workspace. Only her2 algorithm is supported at this moment. Example command to start learning procedure:

python experiment/train.py --num_cpu 1 --env GazeboNICOL-r1-t1-sl-v1 --algorithm her2 --rollout_batch_size 1 --n_epochs 1 --n_train_rollouts 1 --n_test_rollouts 1 --render 0

Environment Parameters for --env GazeboNICOL-r1-t1-sl-v1:

• t{0: joint space (joint control) | 1: task space (inverse kinematics)}

• s{0: right arm | 1: left arm}

• Currently there are no alternative configurations for r1 and v1

Note: You have to build the project and source the catkin workspace, e.g. catkin build && source ~/catkin_ws/devel/setup.bash in the same terminal that you start the ideas command from.