The motion of UR is not smooth when pushT with eval_real_robot.py

[humphreyde]

Thanks for this great work!
I got some problems when I depolyed it on real robot.
First, I use Spacemouse to collection demonstrations with demo_real_robot.py.
But the control is not smooth, and the robot is still moving even though I'm not moving the spacemouse(I monitor the output of spacemouse and there
is nothing output).
Second, After I train the pushT policy with 100 episodes, I run the eval_real_robot.py.
something it works, but the motion is swing and idle.
Is this normal?
I guess it might have something to do with RTDEInterpolationController.
BTW, I use ur5e.

[YahuanShi]

Hi,
I have the same two issues. How do you fix them? Could you give some advice?

[ZC502]

The 'non-smooth' motion you observed on the real UR robot is a classic symptom of Non-Associative Residuals (NAR) embedded in the simulation data.

Most solvers (including MuJoCo) maintain numerical stability by forcing associativity in discrete time steps, which introduces micro-scale physical artifacts. When a Diffusion Policy learns from these 'toxic' trajectories, it fails to capture the true inertial consistency of real hardware.

I’ve developed SIPA to audit these hidden jitters. If you can provide a CSV sample with trajectories, you can use SIPA to diagnose whether these oscillations are structural artifacts generated by the solver's discrete mode:https://github.com/ZC502/SIPA.git