rlenvs_from_cpp is an effort to provide implementations and wrappers of environments for reinforcement learning algorithms to be used by C++ drivers.
Currently, we provide a minimal number of wrappers for some common Gymnasium (former OpenAI-Gym) environments. Namely
FrozenLakewith4x4mapFrozenLakewith8x8mapBlackjackCliffWalkingCartPoleMountainCarTaxiStateAggregationCartPole(implements state aggregation forCartPole)SerialVectorEnvWrappera vector wrapper for various environments
In addition there are wrappers for
GymWalkenvironment from gym_walkgym-pybullet-dronesfrom gym-pybullet-drones
In general, the environments exposed by the library should abide with dm_env specification.
The following snippet shows how to use the FrozenLake and Taxi environments from Gymnasium.
#include "rlenvs/rlenvs_types_v2.h"
#include "rlenvs/envs/gymnasium/toy_text/frozen_lake_env.h"
#include "rlenvs/envs/gymnasium/toy_text/taxi_env.h"
#include <iostream>
#include <string>
#include <unordered_map>
#include <any>
namespace example_1{
const std::string SERVER_URL = "http://0.0.0.0:8001/api";
void test_frozen_lake(){
rlenvs_cpp::envs::gymnasium::FrozenLake<4> env(SERVER_URL);
std::cout<<"Environame URL: "<<env.get_url()<<std::endl;
// make the environment
std::unordered_map<std::string, std::any> options;
options.insert({"is_slippery", true});
env.make("v1", options);
std::cout<<"Is environment created? "<<env.is_created()<<std::endl;
std::cout<<"Is environment alive? "<<env.is_alive()<<std::endl;
std::cout<<"Number of valid actions? "<<env.n_actions()<<std::endl;
std::cout<<"Number of states? "<<env.n_states()<<std::endl;
// reset the environment
auto time_step = env.reset(42);
std::cout<<"Reward on reset: "<<time_step.reward()<<std::endl;
std::cout<<"Observation on reset: "<<time_step.observation()<<std::endl;
std::cout<<"Is terminal state: "<<time_step.done()<<std::endl;
//...print the time_step
std::cout<<time_step<<std::endl;
// take an action in the environment
auto new_time_step = env.step(rlenvs_cpp::envs::gymnasium::FrozenLakeActionsEnum::RIGHT);
std::cout<<new_time_step<<std::endl;
// get the dynamics of the environment for the given state and action
auto state = 0;
auto action = 1;
auto dynamics = env.p(state, action);
std::cout<<"Dynamics for state="<<state<<" and action="<<action<<std::endl;
for(auto item:dynamics){
std::cout<<std::get<0>(item)<<std::endl;
std::cout<<std::get<1>(item)<<std::endl;
std::cout<<std::get<2>(item)<<std::endl;
std::cout<<std::get<3>(item)<<std::endl;
}
// synchronize the environment. environment knows how
// to cast std::any
env.sync(std::unordered_map<std::string, std::any>());
// close the environment
env.close();
}
void test_taxi(){
rlenvs_cpp::envs::gymnasium::Taxi env(SERVER_URL);
std::cout<<"Environame URL: "<<env.get_url()<<std::endl;
// make the environment
std::unordered_map<std::string, std::any> options;
env.make("v3", options);
std::cout<<"Is environment created? "<<env.is_created()<<std::endl;
std::cout<<"Is environment alive? "<<env.is_alive()<<std::endl;
std::cout<<"Number of valid actions? "<<env.n_actions()<<std::endl;
std::cout<<"Number of states? "<<env.n_states()<<std::endl;
// reset the environment
auto time_step = env.reset(42);
std::cout<<"Reward on reset: "<<time_step.reward()<<std::endl;
std::cout<<"Observation on reset: "<<time_step.observation()<<std::endl;
std::cout<<"Is terminal state: "<<time_step.done()<<std::endl;
//...print the time_step
std::cout<<time_step<<std::endl;
// take an action in the environment
auto new_time_step = env.step(rlenvs_cpp::envs::gymnasium::TaxiActionsEnum::RIGHT);
std::cout<<new_time_step<<std::endl;
// get the dynamics of the environment for the given state and action
auto state = 0;
auto action = 1;
auto dynamics = env.p(state, action);
std::cout<<"Dynamics for state="<<state<<" and action="<<action<<std::endl;
for(auto item:dynamics){
std::cout<<std::get<0>(item)<<std::endl;
std::cout<<std::get<1>(item)<<std::endl;
std::cout<<std::get<2>(item)<<std::endl;
std::cout<<std::get<3>(item)<<std::endl;
}
// close the environment
env.close();
}
}
int main(){
std::cout<<"Testing FrozenLake..."<<std::endl;
example_1::test_frozen_lake();
std::cout<<"===================="<<std::endl;
std::cout<<"Testing Taxi..."<<std::endl;
example_1::test_taxi();
std::cout<<"===================="<<std::endl;
return 0;
}
Some algorithms, such as Monte Carlo, require that we should generate a trajectory, example 3 shows how to do this. Various RL algorithms using the environments can be found at cuberl
The general use case is to build the library and link it with your driver code to access its functionality.
Furthermore, the Gymnasium, gym_pybullet_drones environments are accessed via a client/server pattern.
Namely, they are exposed via an API developed using FastAPI.
You need to fire up the server, see dependencies, before using the environments in your code. To do so
cd rest_api && ./start_uvicorn.sh
By default the uvicorn server listents on port 8001. Change this if needed. You can access the OpenAPI specification at
http://0.0.0.0:8001/docs
Note that currently the implementation is not thread/process safe i.e. if multiple threads/processes access the environment a global instance of the environment is manipulated. Thus no session based environment exists.
The library has the following general dependencies
Using the Gymnasium environments requires Gymnasium installed on your machine. In addition, you need to install
In addition, the library also incorporates, see (src/extern), the following libraries
There are extra dependencies if you want to generate the documentation. Namely,
- Doxygen
- Sphinx
- sphinx_rtd_theme
- breathe
- m2r2
The usual CMake based installation process is used. Namely
mkdir build && cd build && cmake ..
make install
You can execute all the tests by running the helper script execute_tests.sh.
It is likely that you are missing the boost_system library with your local Boost installation. This may be the case is you installed boost via a package manager. On a Ubuntu machine the following should resolve the issue
sudo apt-get update -y
sudo apt-get install -y libboost-system-dev