How to deal with a big joint_vel_error?

[AgentEXPL]

Mean episode rew_action_rate: -0.3421 Mean episode rew_alive: 0.3695 Mean episode rew_ang_vel_xy: -0.0022 Mean episode rew_dof_acc: -0.2654 Mean episode rew_dof_pos_limits: -0.0617 Mean episode rew_dof_torque_limits: -0.1109 Mean episode rew_dof_vel: -0.0731 Mean episode rew_orientation: -0.0031 Mean episode rew_pitch: -0.0058 Mean episode rew_pitch_rate: -0.1473 Mean episode rew_roll: -0.0004 Mean episode rew_roll_rate: -0.0726 Mean episode rew_torque_penalty: -0.0219 Mean episode rew_tracking_joint_dof: 1.1962 Mean episode rew_tracking_joint_vel: 0.3092 Mean episode rew_tracking_keybody_pos: 1.2839 Mean episode rew_tracking_keybody_pos_global: 1.3998 Mean episode rew_tracking_root_angular_vel: 0.6716 Mean episode rew_tracking_root_linear_vel: 0.7249 Mean episode rew_tracking_root_pose_delta_local: 1.4779 Mean episode rew_tracking_root_rotation: 0.3456 Mean episode error_tracking_joint_dof: 0.3333 Mean episode error_tracking_joint_vel: 5.5626 Mean episode error_tracking_keybody_pos: 0.0743 Mean episode error_tracking_root_ang_vel: 1.3100 Mean episode error_tracking_root_pose_delta_local: 0.0014 Mean episode error_tracking_root_rotation: 0.0891 Mean episode error_tracking_root_rotation_delta_local: 0.1283 Mean episode error_tracking_root_translation: 0.0317 Mean episode error_tracking_root_vel: 0.3878
The training process seems to be stable, since the reward function exhibits little variation.
The robot has a joint chattering issue. The direct cause of the joint chattering is that the policy applies saturated torque with high gain for position tracking, while the velocity error is large.