koko-mujoco

Requirement

• Mujoco1.55
• OpenAI Gym
• OpenAI Mujoco-py

Repo structure

├── README.md
├── simulate.py
├── train.py
├── assets
│   ├── koko_full.xml
│   └── STL files
└── koko_gym
    └── envs
        ├── assets
        │   ├── koko_reacher.xml
        │   └── STL files
        ├── __init__.py
        └──  koko_reacher.py

Explaination for each file

• koko_full.xml

MJCF file for the Blue robot. Actuated gripper installed. Having following actuators (joints).

    <actuator>
        <motor ctrllimited="true" ctrlrange="-1.0 1.0" gear="2.0" joint="base_roll_joint" />
        <motor ctrllimited="true" ctrlrange="-1.0 1.0" gear="2.0" joint="shoulder_lift_joint" />
        <motor ctrllimited="true" ctrlrange="-1.0 1.0" gear="2.0" joint="shoulder_roll_joint" />
        <motor ctrllimited="true" ctrlrange="-1.0 1.0" gear="2.0" joint="elbow_lift_joint" />
        <motor ctrllimited="true" ctrlrange="-1.0 1.0" gear="2.0" joint="elbow_roll_joint" />
        <motor ctrllimited="true" ctrlrange="-1.0 1.0" gear="2.0" joint="wrist_lift_joint" />
        <motor ctrllimited="true" ctrlrange="-1.0 1.0" gear="2.0" joint="wrist_roll_joint" />
        <position ctrllimited="true" ctrlrange="0 1.05" gear="1.0" joint="robotfinger_actuator_joint" />
        <position ctrllimited="true" kp="1.0" ctrlrange="0 1.4" joint="right_fingerlimb_joint" />
        <position ctrllimited="true" kp="1.0" ctrlrange="-1.4 0" joint="right_fingertip_joint" />
        <position ctrllimited="true" kp="1.0" ctrlrange="0 1.4" joint="left_fingerlimb_joint" />
        <position ctrllimited="true" kp="1.0" ctrlrange="-1.4 0" joint="left_fingertip_joint" />
    </actuator>

Since no URDF <mimic> tag equivalent exists in MJCF, the grippers (last four actuators) are actuated by a position controller that takes the current robotfinger_actuator_joint angle as an input (fingerlimb_joint moves positive and fingertip_joint goes negative to make the tips parallel to each other).

• koko_reacher.py

OpenAI Gym environment for Blue. reacher.step takes 1x8 size action array. The actuator of the gripper joints cannot be controlled respectively but will be controlled at once using robotfinger_actuator_joint's angle as the position input. You can also set your favorite reward signal in a step function.

• koko_reacher.xml

koko_full.xml with target object.

• train.py

Training loop using random controller. Add your favorite algorithm to train the policy.

• simulate.py

Runs the Mujoco-py viewer simulator for 5000 time steps. Use this for test run your trained policy.

Reference

• https://github.com/openai/gym
• https://github.com/openai/mujoco-py
• http://www.mujoco.org/