Learn

This package contains all of the python code used for learning. The code is based on Keras which is based on Tensorflow.

Create conda environment and activate it

Anaconda

If not done yet, install anaconda by following the instructions here. Then reate a anaconda environment, activate it and install the requirements in requirements.txt.

conda create -n <env-name> python=3.6
source activate <env-name>

B.3. Install the required python dependencies

pip install -r requirements.txt

Using The system

Train a simple policy to navigate a particle

python3 trainModel.py --config=settings/particleSim/PPO/PPO.json

Train a model to imitate humanoid motion.

python3 doodad_trainModel.py --config=settings/terrainRLImitate/PPO/Flat_Tensorflow_NoPhase.json

Train a humanoid3d LLC for heirarchical training

python3 doodad_trainModel.py --config=settings/terrainRLImitate3D/PPO/Humanoid_Flat_Tensorflow_MultiAgent_WithObs_LLC_v3.json

Train a hierarchical model to navigate agents across many different scenarios

python3 doodad_trainModel.py --config=settings/terrainRLMultiChar/HLC/TD3/ScenarioMixture_WithObs_SimpleReward_Humanoid_1_Tensorflow_v4.json --log_comet=true --shouldRender=false --bootstrap_samples=1 --run_mode=local_docker --meta_sim_samples=4 --meta_sim_threads=4

Running meta simulations

These simulations are designed to sample a few simulations in order to get a more reasonable average of the performance of a method.

python3 doodad_trainModel.py --config=tests/settings/particleSim/PPO/PPO_KERAS_Tensorflow.json --metaConfig=settings/hyperParamTuning/elementAI.json --meta_sim_samples=5 --meta_sim_threads=5 --tuning_threads=2

Info about how the code works

1. This code uses doodad to help make running many simulations on different compute systems easy.
2. Hyperparameters can be sampled using the --tuningConfig= command on the command line.
3. One of the simulation platforms TerrainRLSim provides a large number of physics-based robot simulations. This library is coded in C++ and very fast.