I can see how integrating the sensed torque to get hte planned motion can be tricky. It's like integrating an acceleration to get the displacement. And all numerical integration is susceptible to "drifting". The tolerance in numerical integration can lead to larger error in the 2nd integral for sure. Also, the sensed torque itself isn't "exact" but within numerical tolerance, so that's another possible source of discrepancy.
The sensed torque you get from the planned motion exercise, I think, is a great way to find out about the required sizing of components, e.g. the ratings of a motor. Then in reality, one would control the inputs to the actuator to get to the desired position. I wonder if you can try something similar: control the torque to get to the desired position at each time interval. Then compare the torque input into the joints to the previous sensed one, as well as the motion trajectory. I would expect with a reasonable controller, the trajectory will be similar, and the force be close but not exact. I think I've seen the forces/torques tends to be more "spikey" as the controller try to follow the desired positions.
