Detailed Explanation of Action

A general description of our robots and actions can be found here. Below is a detailed explanation of the agent's action.

All values in action correspond to the normalized target values of controllers.

In OpenCabinetDoor and OpenCabinetDrawer, where the robot has one arm:

• action[0:2]: x,y velocity of the moving platform, the velocity is in the robot frame, which means this is an ego-centric action. This is controlled by velocity PID
• action[2]: rotate the moving platform, controlled by velocity PID
• action[3]: change the height of the robot body, controlled by velocity PID
• action[4:11]: change joint angles of the robot arm, controlled by velocity PID. The order of the joints can be found here.
• action[11:13]: change the positions of fingers in the gripper, controlled by position PID

In PushChair and MoveBucket, where the robot has two arms:

• action[0:2]: x,y velocity of the moving platform, the velocity is in the robot frame, which means this is an ego-centric action. This is controlled by velocity PID
• action[2]: rotate the moving platform, controlled by velocity PID
• action[3]: change the height of the robot body, controlled by velocity PID
• action[4:11]: change joint angles of the right arm, controlled by velocity PID. The order of the joints can be found here.
• action[11:13]: change the positions of fingers in the right gripper, controlled by position PID
• action[13:20]: change joint angles of the left arm, controlled by velocity PID. The order of the joints can be found here.
• action[20:22]: change the positions of fingers in the left gripper, controlled by position PID