Q&A¶

How to compile my application?¶

We highly recommend using Clion. If you are using Clion, add -DRAI_APP=<your appfile location in applications> in the cmake flag. For example, -DRAI_APP=examples/poleBalwithDDPG.

If you are not using Clion, then:

cd $RAI_ROOT
mkdir build && cd build
cmake .. -DRAI_APP=<your appfile location in applications>
make -j

RAI is an ordinary cmake project. All you have to worry is to set the RAI_APP flag in cmake.

How to add a new application?¶

The best way is to copy one of the example folder to another folder in application and rewrite CMakeLists.txt and the main cpp file. The outermost CMakeLists.txt file will include your application folder once you set RAI_APP flag in cmake.

How to add a new task?¶

Your task inherits from Task class in RAI/include/rai/tasks/Task.hpp. Copy RAI/include/rai/tasks/newTaskTemplate/newTask.hpp to a new task folder and modify.

How to add a new algorithm?¶

Algorithm classes do not inherit any class and you can write in freely. The best practice is to incorporate already existing core modules without reinventing them.

How to add a new neural net?¶

Neural net shapes are called graph structure in RAI. RAI/include/rai/function/tensorflow/pythonProtobufGenerator/graph_structure contains a few examples. Make sure that you collect ‘learnable parameters’ in self.l_param_list and ‘all parameters’ in self.a_param_list. Note that ‘learnable parameters’ is a subset of ‘all parameters’. If you are planning to make a commonly used graph_structure, send a pull request and include the figures as well.

How to add a new specialized function?¶

If your algorithm use something other than Qfunction, Valuefunction and Policy, you can implement a new specialized function. Copy existing one (e.g. ValueFunction) and modify it. Make use that you write input_names and output_names. If you have extra parameters to learn, simply add them to gs.l_param_list and gs.a_param_list (GraphStructure objects) before you call __init__ of the parent class. l_param_list means all learnable parameters and a_param_list is all parameters. RAI makes a clear distinction between them.

Does RAI support multi-threading?¶

RAI supports multi-threaded data collection using parallel experience acquisitor, which is based on openmp [1]. openmp will automatically choose the number of cores to use. If you want to specify it explicitly, call :code:’omp_set_num_threads(nThread)` in the beginning of your app file. Multi-gpu support is not implemented yet.

Does RAI have a cluster-support?¶

Not yet.

Does RAI support OpenAI Gym?¶

Not yet.

Who is developing RAI?¶

Jemin Hwangbo, Robotic System Laboratory, ETH Zurich.

How can I contribute?¶

Send a pull request on the bitbucket repo. Please be specific when you explain the new features or bug fixes.

Known Issues¶

Compatibility with Gazebo¶

Tensorflow install from source also installs a specific version of libprotobuf. This can cause Gazebo to malfunction. The only work around we found is to install Gazebo from source.

[1]	http://www.openmp.org/