Gym Research Setup

This project is needed for quick and convenient testing of hypothesis at the gymnasium (and IsaacGym).

Quick start

Important

Before launch this project make sure that true ports are indicate for TensorBoard and Streamlit in train.yaml.

• If something runs in this ports, then it will be kills
• Config name for this ports are tensorboard_port and streamlit_port

Setting up the environment

Download and unzip the archive into isaacgym/

IsaacGym Dockerfile

Add this command at the end isaacgym/docker/Dockerfile

• RUN cd isaacgym/IsaacGymEnvs && pip install -q -e .
• RUN pip install --no-cache-dir --use-deprecated=legacy-resolver -r requirements.txt

IsaacGym run.sh

Edit your isaacgym/docker/run.sh so it looks something like this:

#!/bin/bash
set -e
set -u


if [ $# -eq 0 ]
then
    echo "running docker without display"
    docker run -it --network=host --gpus=all --name=isaacgym_container isaacgym /bin/bash
else
    export DISPLAY=$DISPLAY
	echo "setting display to $DISPLAY"
	xhost +
	docker run -it --rm -v /tmp/.X11-unix:/tmp/.X11-unix -v $(pwd):/opt/isaacgym -e DISPLAY=$DISPLAY --network=host --gpus=all -p 6006:6006 -p 8501:8501 --name=isaacgym_container isaacgym /bin/bash
	xhost -
fi

IsaacGym build.sh

Edit your isaacgym/docker/build.sh so it looks something like this:

#!/bin/bash
set -e
set -u

if [ ! -d "./isaacgym/IsaacGymEnvs" ]; then
	git clone https://github.com/isaac-sim/IsaacGymEnvs ./isaacgym/IsaacGymEnvs
fi

docker build --network host -t isaacgym -f isaacgym/docker/Dockerfile .

Build env

sh isaacgym/docker/build.sh

• It must to do only once

Connect to env

sh isaacgym/docker/run.sh :0

Run inference in container

python run_inference.py +setup=LunarLander

• You can inference from checkpoint with python run_inference.py +checkpoint_path=outputs/DATE/TIME
  • Where DATE and TIME your path to launch

Run train in container

python run_train.py +setup=LunarLander

• For mo details go to link

Streamlit and TensorBoard

You can open localhost with tensorboard_port and streamlit_port to see that

• checkbox with visible in streamlit is responsible for visible in TensorBoard
• Input with alias is responsible for alias in TensorBoard

More details

How to clear all launchs?

python run_clear.py

How to run train from checkpoint?

python run_train.py +setup=LunarLander +checkpoint_path=outputs/DATE/TIME

• You can drop past information from tensorboard with +checkpoint_drop_past=True
  • If you use this flag, then will be use best model, else last

How to run train without saving model?

python run_train.py +setup=LunarLander +mode=debug