Adds Locomanipulation Environment with G1 for Mimic workflow

[michaellin6]

Description

Contributors: @huihuaNvidia2023, @rwiltz, @life1ess, @peterd-NV, @yanchangNvidia

These changes introduce new environments: Isaac-PickPlace-Locomanipulation-G1-Abs-v0 and Isaac-PickPlace-FixedBaseUpperBodyIK-G1-Abs-v0. These environments, specially the locomanipulation one, feature:

1. The G1 robot with the retargeter for 3-finger dexterous hand to operator hands tracked by Apple Vision Pro
2. A locomotion policy trained with RL that enables G1 to walk and follow a input velocity command (vx, vy, wz, h) linear velocity in two axes, angular velocity around its pelvis' z-axis, and height distance of the pelvis from the feet.
3. Integration with upper body controller that is driven by Inverse Kinematics.

Below is a diagram showing the high level controller composition. The environments introduced have the waist controlled by IK, but it is possible to configure it such that waist DOFs are controlled by the lower body policy as well.

This code has been tested through the Mimic pipeline, and the output dataset was used to train a robomimic policy that achieves a success rate of 92% from 50 policy rollouts.

Fixes # (issue)

Some fixes included in these changes:

1. Previously, Inverse kinematics relied on an input URDF of the robot to load a pinocchio kinematics model. This URDF was exported from the robot asset USD through USD to URDF export extension, however, if the extension is not setup properly, the joint ordering can change causing unexpected behavior due to mismatch of joint ordering between the IK controller and that of Isaac Lab. For the G1 asset, we are for now using a URDF asset directly instead of relying on the URDF exporter.
2. Previously, the Inverse Kinematics controller assumed that for its kinematics model, only the joints that are being controlled by IK need to be updated, but when combined with other controllers on other joints this assumption breaks. Added a new kinematics configuration class:PinkKinematicsConfiguration that allows for all joints of an articulation to be updated while only a subset are controlled by IK.

Type of change

• New feature (non-breaking change which adds functionality)

Screenshots

Robomimic policy for PickPlace trained with the Locomanipulation environment. We achieve 92% success rate from 50 rollouts.

Checklist

• I have run the pre-commit checks with ./isaaclab.sh --format
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• I have updated the changelog and the corresponding version in the extension's config/extension.toml file
• I have added my name to the CONTRIBUTORS.md or my name already exists there

⚠️ This PR depends on #3149 and should only be merged after that PR

[rwiltz]

LGTM from the teleop side of things.

[huihuaNvidia2023]

Done the first pass. Will look more into the tests.

[kellyguo11]

can we upload agile_locomotion.pt to nucleus instead of hosting it in github?

[michaellin6]

Just rebased this branch to ToT. Will address all the previous review comments shortly.

[huihuaNvidia2023]

LTGM with two minor comments. Happy to discuss alternatives.

[Cdfghglz]

Hi, will the agile_locomotion training code be published in any form?

[ooctipus]

Implementation looks good to me,
Most concern were legacy issues regarding where to put robot - tasks specific files and specification on gym module registration. It might not be scope of this PR, but will be good to start address them earily

[kellyguo11]

Seeing timeout in test_environments.py and failure in test_environments_with_stage_in_memory.py in CI. could you check locally if the tests are failing or whether we need to increase the timeouts for them?