Hello, thank you for sharing your work.
We are currently trying to reproduce the results of your paper using a separate dataset focused around soccer skills/drills.
We were able to successfully make the motion tracker work in MuJoCo, but it completely freaks out and doesn't work when deployed on the real robot.
We've attempted adding domain randomization, performing system identification, reducing action scale during training and deployment, etc. to no avail.
Reducing action scale during training makes it so training doesn't converge and reducing action scale during deployment allows the robot to follow the trajectory well but is unable to keep balance during it.
We've maybe thought it's because we haven't trained a default/standing pose to the robot as it's not included in our dataset, but was wondering if there are any other protocols that you guys have taken to remedy the sim2real gap.
Accordingly, we've tried putting the robot in a balanced state through the Unitree provided commands and transitioning into the reference motion trajectories, but showed the exact same symptoms.
We are aware that you guys are awaiting the company's approval for the full release of the code, but any assistance regarding this matter would be greatly helpful.
Thank you.
Hello, thank you for sharing your work.
We are currently trying to reproduce the results of your paper using a separate dataset focused around soccer skills/drills.
We were able to successfully make the motion tracker work in MuJoCo, but it completely freaks out and doesn't work when deployed on the real robot.
We've attempted adding domain randomization, performing system identification, reducing action scale during training and deployment, etc. to no avail.
Reducing action scale during training makes it so training doesn't converge and reducing action scale during deployment allows the robot to follow the trajectory well but is unable to keep balance during it.
We've maybe thought it's because we haven't trained a default/standing pose to the robot as it's not included in our dataset, but was wondering if there are any other protocols that you guys have taken to remedy the sim2real gap.
Accordingly, we've tried putting the robot in a balanced state through the Unitree provided commands and transitioning into the reference motion trajectories, but showed the exact same symptoms.
We are aware that you guys are awaiting the company's approval for the full release of the code, but any assistance regarding this matter would be greatly helpful.
Thank you.