This is a gym environment, using ur5 robot and robotiq 3 finger gripper, in mujoco simulator.

The robot xml file is from http://www.mujoco.org/forum/index.php?resources/universal-robots-ur5-robotiq-s-model-3-finger-gripper.22/

This package includes:

• Robot environments
  • ur5 env
  • gripper env
• Task environments
  • UR5GripperCatchBall-v0

• gym-ur5_gripper
  • Overview
  • install
  • Robot Environments
    • UR5 Env
  • Task Environments
    • UR5 Reach Env
      • Observation space
      • Reward function
    • Gripper

$ git clone https://github.com/wangcongrobot/gym-ur5_gripper.git
$ cd gym-ur5_gripper
$ virtualenv env --python=python3
$ pip install -e .
$ python gym-ur5_gripper/tests/test_ur5_gripper_env.py

• Action space:

  • Cartesian space: hand 3D pose (x,y,z)
  • Gripper action: open/close

• Observation space:

  • arm joint positions
  • end-effector pose
  • end-effector velocity

• UR5's observation space
• plus object position and orientation

In this environment, the reward function is given by:

• the distance between the end-effector and the desired position
• plus a bonus when the end-effector is close to the desired position

Here is the code used to compute the reward function:

The Robotiq 3 finger gripper has 11 dof, the control mode includes torque control, position control and so on.

We use position control, and change the 11-dof joint control into a 1-dof open/close action.

def gripper_format_action(self, action):
    """ Given (-1,1) abstract control as np-array return the (-1,1) control signals
    for underlying actuators as 1-d np array
    Args:
        action: 1 => open, -1 => closed
    """
    movement = np.array([0, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1])
    return -1 * movement * action

P.S.: 1 => open, 0 => close, (0,1) => grasp