APPLR

Jackal Navigation Environment

Two OpenAI gym environments with continuous and discrete actions spaces can be found under jackal_navi_envs. They are registered as jackal_discrete-v0 and jackal_continuous-v0. To use the environment locally, you need a workspace with all the jackal dependencies installed. Assume you already set up the Jackal simulation, one extra package jackal_helper need to be installed.

• Under the parent folder of your workspace:

cd src
git clone https://github.com/Daffan/jackal_helper.git
cd ..
source devel/setup.bash
catkin_make

• The following python dependency need to be installed:

pip3 install gym defusedxml pyyaml pyquaternion rospkg

• To test the installation:

source <your_workspace>/devel/setup/bash
python3 scripts/test_env.py

Above command will run a Jackal_navigation_env 10 episodes.

Singularity container

• build the singularity image

sudo singularity build --notest APPLR_melodic.simg APPLR_melodic.def

• Run test in container

./singularity_run.sh python3 scripts/test_env.py

Train td3 policy on HTCondor

• Check all the environment and training related configuration:

cat continuous/config/td3_condor.json

• Run the central learning node locally at host

./executable/run_central_node.sh

Running the central node locally is currently recommanded. A crowded cluster will usually idle your jobs. When you jobs is idled and recovered, it will initialize a new policy! Run it locally can prevent the issue.

• Run all the actors nodes:

pytho3n gen_sub1.py --num_env <num_env in your config file>

Test your policy on HTCondor

• Run test

python3 scripts/test_policy_condor.py --model <path/to/logfolder> --policy <policy_file_name> --test

• Get the test report

python3 scripts/report_test.py

This will print the averaging of some metrics and generate report.json with all test results.