Add QuadrupedAMP task, based on HumanoidAMP method

[dtch1997]

Continuation of #111

Hello, thank you very much for open-sourcing IsaacGymEnvs.

As part of my research, I have created a draft implementation of AMP for legged quadrupeds.

To test pre-trained model:

cd isaacgymenvs
python train.py task=QuadrupedAMP task.env.motionFile=data/motions/quadruped/dog_trot_hq.txt test=True checkpoint=data/checkpoints/QuadrupedAMP/QuadrupedGait/dog_trot_hq.pth num_envs=32 headless=False

To train a new model:

python train.py task=QuadrupedAMP task.env.motionFile=data/motions/quadruped/dog_trot_hq.txt

Sample trotting example (dog_trot_hq.txt):

trot_viz_ref.mp4

At the moment, there is some unexplained drift in the policy. I am unsure of the cause but I am hopeful that upgrading to better motion-capture will fix this issue.

Motion-capture obtained by:
(i) copies of https://github.com/erwincoumans/motion_imitation/tree/master/motion_imitation/data/motions
(ii) taking trained policies from motion_imitation repo and recording the robot state at every env step. (dog_trot_hq.txt, dog_spin_hq.txt, dog_pace_hq.txt`)

Let me know what you think :)

[ViktorM]

It's very impressive @dtch1997, great work! Need to discuss with @gavrielstate the next steps.

[dtch1997]

Thanks so much for the kind words @ViktorM ! I look forward to hearing about the next steps.

It is not included in this PR yet but I have much better results now, which can be viewed here: https://drive.google.com/drive/folders/1w0VEceadkjB0cewBJOzVZmT5wg9pLACt?usp=sharing