How to Validate a “Good” Delta-Action?

[zhang705]

I have deployed a pre-trained policy and collected motion actions, but when I bring these motion actions into the simulation environment, compute joint torques with motion_action + delta_action, the robot usually falls immediately and cannot maintain for the set episode_length_s until the last iteration.
So, how can I determine if I have trained an appropriate delta action policy? Is the correct criterion that the robot's motion action + delta action should enable robot maintain the ref motion for the set episode_lenght_s?