Go1 Legged Gym Tutorial

0. Abbreviations

• LL or LLP - Low-level policy, which converts directional commands (linear velocity on the x-y plane and angular velocity around the UP axis) into joint angles

1. Install everything

Create a new conda environment and follow the install instructions from the 2 readmes in the subdirectories:
https://github.com/simonchamorro/legged-gym-rl

• You should use Python 3.8.
• Due to the isaacgym depency, this only works on a machine with a CUDA GPU and at least Nvidia Pascal arch (GTX 1080 or later).
• You need a specific version of pytorch, as mentioned in the readme.
• Because of this, I highly recommend making a new conda env and not trying to install in an existing env.
• You need a specific IsaacGym version. Ideally, Preview-3 but Preview-4 shouuuuuld work too. P3 can be found here (personal backup): https://drive.google.com/file/d/1ksA2q4aKpgROFClEEAzy9tNGblr_tzFt/view?usp=share_link

2. Training a low-level policy

We assume that you're doing this all in the legged_gym/legged_gym subdirectory where you should see the scripts folder (https://github.com/simonchamorro/legged-gym-rl/tree/main/legged_gym/legged_gym):

​​​​cd legged_gym/legged_gym

To check that everything works correctly, try the command:

​​​​python scripts/train.py --task=go1_mrss_novel

If you want to see the robots, you can attach the sim viewer (only works if a physical screen is attached to the GPU):

​​​​python scripts/train.py --task=go1_mrss_novel --use_viewer

If you're running this on a cluster machine, make sure to use the headless flag to disable all rendering facilities:

​​​​python scripts/train.py --task=go1_mrss_novel --headless

When training is done, the trained policy can be found under the name of the environment in the logs/ folder.

2.1 Additional parameters

• --seed 1234 you can set a fixed seed, which is always advisible for RL envs
• --experiment_name ladeeda I think by default all experiments are logged to Weights & Biases and this helps identify the experiment
• --task X… there are like 10 million different small variations on environments and a full list can be found here (bottom of the file: pupperfetch/legged_gym/envs/__init__.py

3. Playing back a low-level policy

To visualize a low-level policy with random commands, use the 02-play script. There are also better script that test specific features or give you a remote control for the LL policy.

3.1 Low-level policy with random commands

​​​​python scripts/02-play.py --task go1_flat --load_run 23Jan16_12-53-01_v1-0.5 --use_viewer

Replace --task go1_flat with whatever task and replace --load_run 23J... with the name of your trained policy (can be found under /logs/TASK_NAME/RUN_NAME) and add the --use_viewer flag to actually open a window and visualize the lil puppers.