Initial arm_hand state in GraspSim

[Ralph-cong]

Hi~ Thanks for the work. During the training of the grasp skill, I found that the arm_hand is initialized with a group of predefined action list arm_hand_prepare_dof_pos_list, what does that mean? And the issue is that the hand is far above the blocks，which makes it impossible to complete the grasp task. Should the arm_hand be initialized with the final arm_hand state like what the insert_sim do?

` self.arm_hand_default_dof_pos = torch.zeros(self.num_arm_hand_dofs, dtype=torch.float, device=self.device) self.arm_hand_default_dof_pos[:7] = torch.tensor([0.0621, 0.3445, 0.3000, -2.3935, -0.2406, 2.7193, 2.1295], dtype=torch.float, device=self.device)

    self.arm_hand_default_dof_pos[7:] = to_torch([0.0, -0.174, 0.785, 0.785,
                                        0.0, -0.174, 0.785, 0.785, 0.0, -0.174, 0.785, 0.785, 0.0, -0.174, 0.785, 0.785], dtype=torch.float, device=self.device)
    
    self.arm_hand_default_dof_pos[7:] = scale(torch.ones(16, dtype=torch.float, device=self.device), 
                                            self.arm_hand_dof_lower_limits[7:23], self.arm_hand_dof_upper_limits[7:23])


    self.arm_hand_prepare_dof_poses = torch.zeros((self.num_envs, self.num_arm_hand_dofs), dtype=torch.float, device=self.device)
    self.end_effector_rotation = torch.zeros((self.num_envs, 4), dtype=torch.float, device=self.device)

    self.arm_hand_prepare_dof_pos_list = []
    self.end_effector_rot_list = []
    self.arm_hand_insertion_prepare_dof_pos_list = []

    # rot = [0, 0.707, 0, 0.707]
    self.arm_hand_prepare_dof_pos = to_torch([0.0, -0.49826458111314524, -0.01990020486871322, -2.4732269941140346, -0.01307073642274261, 2.00396583422025, 1.5480939705504309,
                                            0.0, -0.174, 0.785, 0.785,
                                        0.0, -0.174, 0.785, 0.785, 0.0, -0.174, 0.785, 0.785, 0.0, -0.174, 0.785, 0.785], dtype=torch.float, device=self.device)

... `