Swift4RL is an open source organization that explores new way of working with Reinforcement Learning combining Swift for TensorFlow (S4TF), Unity ML-Agents and possibly other environments in the future. Python has been the main programming language for data scientist for so long eventhough it has some drawbacks such as being slow, concurrency (GIL), and relying on native bindings for wiring Python with C++ and CUDA for resource intensive applications. On the other hand Swift for Tensorflow provides easy way to write custom ops, debugging, aesthetics and design which allows power users like AI researches to become productive. Another advantage of Swift is that it's been built for mobile devices. Therefore it is pretty lightweight. Wouldn't it be awesome to train baseline models by taking advantage of Unity's physic engine and deploy them on the real world mobile robots.

• Xcode 12
• macOS 10.15.6 and above
• Swift 5.3
• Unity Hub
• protoc-4.0.0-rc-2
• grpc_csharp_plugin

Install rake and xcodeproj as follows and then execute rake dependencies that will generate xcode project with deployment target macOS 10.15

[sudo] gem install rake
[sudo] gem install xcodeproj
rake dependencies

To run with python gym environments one must have a the python packages installed. Swift must also know where python is where the dependencies are installed.

Install python 3.8 as a system python version, not a virtual env. We could not get that to work, e.g. in a venv or anaconda env.

brew install [email protected]

We need python dependencies for only observing training in tensorboard, other than that we use Swift for Tensorflow in the project. In your newly installed python 3.8 install dependencies. The path is on my mac "/usr/local/Frameworks/Python.framework/Versions/bin/python3"

Use your path as from the example above

PATH/python3 -m pip install tensorflow tensorboard

If you have not installed pip in the python version above do it as in here. With the python path as above.

Swift makes it possible to choose which python version to use. Documentation is here, but not the easiest to get it to work. We could not get it to work with a venv or anaconda env.

You will need to change the path to the environment inside Run/main.swift to an executable environment. Docs to create an executable environment link

PYTHON_LIBRARY=PATH/python3 swift run

Then the model should run and train a PPO model on the environment. This also saves logs to tensorboard via the python api. To run the tensorboard with correct logdir run:

tensorboard --logdir=logs