How to ensure identical TCP twist (linear & angular velocity) between no-load and loaded robot in IsaacLab, under the same trajectory?

[YeJiAce]

Hello, I am working on dynamic parameter identification in IsaacLab and have an issue:

- I send exactly the same joint position commands to the robot
- I run two trials: no-load and grasping an object (loaded)
- Actual joint positions joint_pos are nearly identical in both cases
- However, TCP linear and angular velocities (body_link_vel_w) show small occasional deviations between the no-load and loaded cases, even though their overall trends are nearly identical.

I understand this is because:Under implicit PD position control, payload changes inertia → different dynamic response → different actual joint velocities and twist.

Since the twist exhibits small deviations, the inertial forces also differ slightly. As a result, I cannot reliably use wrench_loaded - wrench_no_load to obtain clean interaction forces, as this direct subtraction is not fully physically rigorous.

Goal
I hope to make the robot’s TCP twist as consistent as possible under the same joint trajectory, even with different end-effector payloads.

What I have checked
Same joint position commands in both cases
Same robot model, stiffness, damping, effort_limit, velocity_limit
Same physics settings (sim_dt, substeps, solver iterations)
Velocities read in world frame (body_link_vel_w)

Questions

1. I’m curious: what might be a robust way to improve velocity consistency across different payloads in IsaacLab?
2. Would switching to joint velocity control (set_joint_velocity_target) help reduce these small twist deviations?
3. Alternatively, what is a good way to extract the pure contact force exerted by the object from the force-torque sensor readings?

Thanks for your response.

Isaac Lab Version: [2.2.0]
Isaac Sim Version: [5.0.0]