DataLab#

1.  Parameterized human model from Xsens#

For this part, we can extract parameterized URDF human model from Xsens based on a template URDF model.

  1. The template URDF model can be found in template urdf model. In this urdf, we decouple the ball joints of human into 3 rotation joints in the order of Z-X-Y.

  2. By following this example, we can extract the parameters that the template model needs from Xsens mvnx file and generate a new URDF model.

  3. I also have tried to use the joint value in Xsens mvnx file to control this urdf. The trail can be found in this script. However, the result is not good, the motion is quite weird. I analyzed that the reason should be the joint values in Xsens mvnx file rotate based on a different axis with our urdf. In order to leverage the experience in AMP and ASE, I finally chose to use the motion retargeting method provided by them.

2.  Motion retargeting from Xsens#

The motion retargeting is not based on URDF model now, it is built on poselib and MJCF humanoid model.

  1. The poselib can convert the Xsens .fbx file to a .npy file, please check this.

  2. The .npy motion can be visualized in Isaac Gym by following command

    python examples/learning_rl/example_HumanoidASE_RofuncRL.py --task HumanoidViewMotion --motion_file /path/to/motion.npy --headless=False --inference

    Some example data is provided in examples/data/ase and examples/data/amp.