Lifting cube sim2real?

[jiwonyou0420]

Hi, I'm trying to do sim2real on lift cube task with UR3 + custom UMI gripper.

I trained lift cube policy with RL in sim, exported to torch jit(using rsl rl) then deployed on UR3 + UMI gripper.

Due to safety reasons(obviously) UR3 moves much slower than the one in simulator. However, this means that the joint velocities in observation would be different and could cause additional sim2real problems. In fact, when I take a well-trained lift policy and run inference with slow joint velocities(scaled down by 0.001) it shows strange behaviors.

Inference with original velocity
https://github.com/user-attachments/assets/e4860e12-600f-4d70-8dd2-3818cca5fa06

Inference with scaled down velocity
https://github.com/user-attachments/assets/bce07e50-1864-4cbb-ad94-b17d5d976831

Is this a known issue in sim2real? If so, what would be the best practice? Slow down arm in simulation? Do not use joint velocity as observation during training?

Thanks in advance