Skip to content

README_TRAIN.md

Latest commit

 

History

History
78 lines (50 loc) · 2.88 KB

README_TRAIN.md

File metadata and controls

78 lines (50 loc) · 2.88 KB

Installation Train

  1. Install Isaac Lab by following the installation guide. We recommend using the conda installation as it simplifies calling Python scripts from the terminal.

  2. Install git for very large file

sudo apt install git-lfs

  1. Clone the repository separately from the Isaac Lab installation (i.e. outside the IsaacLab directory)

  2. Using a python interpreter that has Isaac Lab installed, install the library

python -m pip install -e source/basic_locomotion_isaaclab

  1. If you want to play with Morphologycal Symmetries, install the repo morphosymm-rl

  2. If you want to play with Adversarial Motion Priors, install the repo amp-rsl-rl from the AMI research lab.

Run a train/play in IsaacLab

  • To train:
python scripts/rsl_rl/train.py --task=Locomotion-Aliengo-Flat --num_envs=4096 --headless
python scripts/rsl_rl/train.py --task=Locomotion-Aliengo-Rough-Blind --num_envs=4096 --headless
  • To test the policy, you can press:
python scripts/rsl_rl/play.py --task=Locomotion-Aliengo-Flat --num_envs=16
python scripts/rsl_rl/play.py --task=Locomotion-Aliengo-Rough-Blind --num_envs=16

Use AMP, Morphological Symmetries, DAGGER or Depth to Heightmap

Each of these modules as a specific README in its own script folder.

Run Hyperparameter Search

echo "import ray; ray.init(); import time; [time.sleep(10) for _ in iter(int, 1)]" | python3 (TERMINAL 1)
python3 ../basic_locomotion_isaaclab/exts/basic_locomotion_isaaclab/basic_locomotion_isaaclab/hyperparameter_tuning/tuner.py --run_mode local --cfg_file ../basic_locomotion_isaaclab/exts/basic_locomotion_isaaclab/basic_locomotion_isaaclab/hyperparameter_tuning/locomotion_aliengo_cfg.py --cfg_class LocomotionAliengoFlatTuner (TERMINAL 2)

Convert XML to USD

We use model from gym-quadruped.

./isaaclab.sh -p scripts/tools/convert_mjcf.py   ../basic_locomotion_isaaclab/scripts/sim_to_sim_mujoco/gym-quadruped/gym_quadruped/robot_model/aliengo/aliengo.xml   ../aliengo.usd   --import-sites   --make-instanceable

Remember to set in the application above, "set as default prim" to the root of the robot. Furthermore, for now, add the following lines in the xml of your robots to make the feet seen as body

<body name="FL_foot" pos="0 0 -0.25">
    <!-- FL_foot only collision -->
    <geom name="FL" class="collision" size="0.0265" pos="0 0 0" />
</body>

Good to know

If you have speed problem in training, may be due to cylinder collision. Then add, while calling train.py or play.py, this arg:

--kit_args="--/physics/collisionApproximateCylinders=true"