Custom Environment

You can find a repo with a corresponding example/template at https://github.com/rl-tools/example

As always we first include the elementary operations

#define RL_TOOLS_BACKEND_ENABLE_OPENBLAS
#include <rl_tools/operations/cpu_mux.h>
#include <rl_tools/nn/operations_cpu_mux.h>
#include <rl_tools/nn_models/operations_cpu.h>
namespace rlt = rl_tools;
#pragma cling load("openblas")

Next we define the datastructures for our new environment. The main data structures are for the state of the environmnet (MyPendulumState) and for the environment itself (MyPendulum). As usual in RLtools, we assemble all template parameters of the environment into a specification (MyPendulumSpecification) so that we do not need to repeat them in every function template. Furthermore we separate out the parameters. With RLtools environments we distinguish between three levels of “state”: - Environment: Compile-time, should not change at runtime (though this is is not enforced to allow for hackability) - Parameters: Constant throughout an episode. This allows for e.g. domain randomization. It can also carry cues for the visualization (the constness during an episode is also not enforce but considered a best practice) - State: Sampled from the initial distribution at the beginning of an episode and can then change on every step

To work with the RLtools API, the environment data structure (MyPendulum) needs to have the fields: - T: Floating point type - TI: Index/unsigned integer type - Parameters: Parameters datastructure (can also purly contain compile-time constexpr), should be a Plain Old Data (POD) structure so that it works well on GPUs and microcontrollers - State: State datastructure, should be a Plain Old Data (POD) for the same reasons - OBSERVATION_DIM: Dimension of the observations - ACTION_DIM: Dimension of the actions

template <typename T>
struct MyPendulumParameters {
    constexpr static T G = 10;
    constexpr static T MAX_SPEED = 8;
    constexpr static T MAX_TORQUE = 2;
    constexpr static T DT = 0.05;
    constexpr static T M = 1;
    constexpr static T L = 1;
    constexpr static T INITIAL_STATE_MIN_ANGLE = -rlt::math::PI<T>;
    constexpr static T INITIAL_STATE_MAX_ANGLE = rlt::math::PI<T>;
    constexpr static T INITIAL_STATE_MIN_SPEED = -1;
    constexpr static T INITIAL_STATE_MAX_SPEED = 1;
};

template <typename T_T, typename T_TI, typename T_PARAMETERS = MyPendulumParameters<T_T>>
struct MyPendulumSpecification{
    using T = T_T;
    using TI = T_TI;
    using PARAMETERS = T_PARAMETERS;
};

template <typename T, typename TI>
struct MyPendulumState{
    static constexpr TI DIM = 2;
    T theta;
    T theta_dot;
};
template <typename TI>
struct MyPendulumFourierObservation{
    static constexpr TI DIM = 3; // cos(theta), sin(theta), theta_dot
};

template <typename T_SPEC>
struct MyPendulum{
    using SPEC = T_SPEC;
    using T = typename SPEC::T;
    using TI = typename SPEC::TI;
    using Parameters = typename SPEC::PARAMETERS;
    using State = MyPendulumState<T, TI>;
    using Observation = MyPendulumFourierObservation<TI>;
    using ObservationPrivileged = Observation;
    static constexpr TI ACTION_DIM = 1;
    static constexpr TI N_AGENTS = 1;
};

Next we can start defining operations on these datastructures. Note that they should be in the rl_tools namespace so that the RLtools algorithms (such as the on-/off-policy runner) can find and dispatch to them. If you want to use functions outside the rl_tools namespace you can just implement proxy functions that call your arbitrary functions. In our case we do not need dynamic memory allocation or initialization, hence just implement them as a NOP. The sample_initial_state function samples random initial states and the initial_state provides a deterministic initial state (for deterministic evaluations). In the case of the pendulum a reasonable choice for the latter could be e.g. the state where it is hanging downwards with zero velocity.

namespace rl_tools{
    template<typename DEVICE, typename SPEC>
    static void malloc(DEVICE& device, const MyPendulum<SPEC>& env){}
    template<typename DEVICE, typename SPEC>
    static void free(DEVICE& device, const MyPendulum<SPEC>& env){}
    template<typename DEVICE, typename SPEC>
    static void init(DEVICE& device, const MyPendulum<SPEC>& env, typename MyPendulum<SPEC>::Parameters& parameters){}
    template<typename DEVICE, typename SPEC, typename RNG>
    static void sample_initial_parameters(DEVICE& device, const MyPendulum<SPEC>& env, typename MyPendulum<SPEC>::Parameters& parameters, RNG& rng){ }
    template<typename DEVICE, typename SPEC>
    static void initial_parameters(DEVICE& device, const MyPendulum<SPEC>& env, typename MyPendulum<SPEC>::Parameters& parameters){ }
    template<typename DEVICE, typename SPEC, typename RNG>
    static void sample_initial_state(DEVICE& device, const MyPendulum<SPEC>& env, const typename MyPendulum<SPEC>::Parameters& parameters, typename MyPendulum<SPEC>::State& state, RNG& rng){
        state.theta     = random::uniform_real_distribution(typename DEVICE::SPEC::RANDOM(), SPEC::PARAMETERS::INITIAL_STATE_MIN_ANGLE, SPEC::PARAMETERS::INITIAL_STATE_MAX_ANGLE, rng);
        state.theta_dot = random::uniform_real_distribution(typename DEVICE::SPEC::RANDOM(), SPEC::PARAMETERS::INITIAL_STATE_MIN_SPEED, SPEC::PARAMETERS::INITIAL_STATE_MAX_SPEED, rng);
    }
    template<typename DEVICE, typename SPEC>
    static void initial_state(DEVICE& device, const MyPendulum<SPEC>& env, const typename MyPendulum<SPEC>::Parameters& parameters, typename MyPendulum<SPEC>::State& state){
        state.theta = -rlt::math::PI<typename SPEC::T>;
        state.theta_dot = 0;
    }
}

In the following we define some helper functions. Note: the usage of rlt::math::xxx for math functions seems tedious over e.g. std::xxx but it allows to dispatch to the right math implementations on GPUs and microcontrollers and hence running the same code on any device.

template <typename T>
T clip(T x, T min, T max){
    x = x < min ? min : (x > max ? max : x);
    return x;
}
template <typename DEVICE, typename T>
T f_mod_python(const DEVICE& dev, T a, T b){
    return a - b * rlt::math::floor(dev, a / b);
}

template <typename DEVICE, typename T>
T angle_normalize(const DEVICE& dev, T x){
    return f_mod_python(dev, (x + rlt::math::PI<T>), (2 * rlt::math::PI<T>)) - rlt::math::PI<T>;
}

Next we implement the most important operations (which resemble the OpenAI gym interface): - step: Takes a state, executes an action and sets the next_state - reward: Returns the reward based on the state, action, and next_state - observe: Observes the state. For fully observed environments this should basically just flatten the ::State data structure and possibly apply some observation noise. For partially observable environments the observation can also just contain parts of the information in the ::State - terminated: Returns a boolean flag signalling if the state is a terminal state

namespace rl_tools{
    template<typename DEVICE, typename SPEC, typename ACTION_SPEC, typename RNG>
    typename SPEC::T step(DEVICE& device, const MyPendulum<SPEC>& env, const typename MyPendulum<SPEC>::Parameters& parameters, const typename MyPendulum<SPEC>::State& state, const Matrix<ACTION_SPEC>& action, typename MyPendulum<SPEC>::State& next_state, RNG& rng) {
        static_assert(ACTION_SPEC::ROWS == 1);
        static_assert(ACTION_SPEC::COLS == 1);
        typedef typename SPEC::T T;
        typedef typename SPEC::PARAMETERS PARAMS;
        T u_normalised = get(action, 0, 0);
        T u = PARAMS::MAX_TORQUE * u_normalised;
        T g = PARAMS::G;
        T m = PARAMS::M;
        T l = PARAMS::L;
        T dt = PARAMS::DT;

        u = clip(u, -PARAMS::MAX_TORQUE, PARAMS::MAX_TORQUE);

        T newthdot = state.theta_dot + (3 * g / (2 * l) * rlt::math::sin(device.math, state.theta) + 3.0 / (m * l * l) * u) * dt;
        newthdot = clip(newthdot, -PARAMS::MAX_SPEED, PARAMS::MAX_SPEED);
        T newth = state.theta + newthdot * dt;

        next_state.theta = newth;
        next_state.theta_dot = newthdot;
        return SPEC::PARAMETERS::DT;
    }
    template<typename DEVICE, typename SPEC, typename ACTION_SPEC, typename RNG>
    static typename SPEC::T reward(DEVICE& device, const MyPendulum<SPEC>& env, const typename MyPendulum<SPEC>::Parameters& parameters, const typename MyPendulum<SPEC>::State& state, const Matrix<ACTION_SPEC>& action, const typename MyPendulum<SPEC>::State& next_state, RNG& rng){
        typedef typename SPEC::T T;
        T angle_norm = angle_normalize(device.math, state.theta);
        T u_normalised = get(action, 0, 0);
        T u = SPEC::PARAMETERS::MAX_TORQUE * u_normalised;
        T costs = angle_norm * angle_norm + 0.1 * state.theta_dot * state.theta_dot + 0.001 * (u * u);
        return -costs;
    }

    template<typename DEVICE, typename SPEC, typename OBS_TYPE_SPEC, typename OBS_SPEC, typename RNG>
    static void observe(DEVICE& device, const MyPendulum<SPEC>& env, const typename MyPendulum<SPEC>::Parameters& parameters, const typename MyPendulum<SPEC>::State& state, const MyPendulumFourierObservation<OBS_TYPE_SPEC>&, Matrix<OBS_SPEC>& observation, RNG& rng){
        static_assert(OBS_SPEC::ROWS == 1);
        static_assert(OBS_SPEC::COLS == 3);
        typedef typename SPEC::T T;
        set(observation, 0, 0, rlt::math::cos(device.math, state.theta));
        set(observation, 0, 1, rlt::math::sin(device.math, state.theta));
        set(observation, 0, 2, state.theta_dot);
    }
    template<typename DEVICE, typename SPEC, typename RNG>
    RL_TOOLS_FUNCTION_PLACEMENT static bool terminated(DEVICE& device, const MyPendulum<SPEC>& env, const typename MyPendulum<SPEC>::Parameters& parameters, const typename MyPendulum<SPEC>::State state, RNG& rng){
        using T = typename SPEC::T;
        return false;
    }
}

Since the training functions for the RL algorithms need to execute these operations they need to be defined before the RL data-collection and training operations. Hence in the following we include the RL (Loop Interface) operations. Note: when setting up your project you might want to assemble the previous data-structure definitions and operations into a header file so that all the #include directives are at the beginning of your code (still remember to include the header files for your environment before the RL operations). A recommended structure (that RLtools follows internally as well) is to put the the environment (for this example) into a my_pendulum directory. Then the datastructures are in my_pendulum/my_pendulum.h and the operations are in my_pendulum/operations_generic.h. operations_generic.h means that these are pure C++, dependency-free operations that can run on any device. If you need to use external libraries (e.g. std::xxx or nlohmann::json) you should separate out these operations into a device specific header, e.g. my_pendulum/operations_cpu.h.

#include <rl_tools/rl/algorithms/ppo/loop/core/config.h>
#include <rl_tools/rl/algorithms/ppo/loop/core/operations_generic.h>
#include <rl_tools/rl/loop/steps/evaluation/config.h>
#include <rl_tools/rl/loop/steps/evaluation/operations_generic.h>

Finally, we can use our new environment and train it using the Loop Interface (same as in the previous chapter):

using DEVICE = rlt::devices::DEVICE_FACTORY<>;
using RNG = decltype(rlt::random::default_engine(typename DEVICE::SPEC::RANDOM{}));
using T = float;
using TI = typename DEVICE::index_t;

using PENDULUM_SPEC = MyPendulumSpecification<T, TI, MyPendulumParameters<T>>;
using ENVIRONMENT = MyPendulum<PENDULUM_SPEC>;

struct LOOP_CORE_PARAMETERS: rlt::rl::algorithms::ppo::loop::core::DefaultParameters<T, TI, ENVIRONMENT>{
    static constexpr TI EPISODE_STEP_LIMIT = 200;
    static constexpr TI TOTAL_STEP_LIMIT = 300000;
    static constexpr TI STEP_LIMIT = TOTAL_STEP_LIMIT/(ON_POLICY_RUNNER_STEPS_PER_ENV * N_ENVIRONMENTS) + 1; // number of PPO steps
};
using LOOP_CORE_CONFIG = rlt::rl::algorithms::ppo::loop::core::Config<T, TI, RNG, ENVIRONMENT, LOOP_CORE_PARAMETERS>;

template <typename NEXT>
struct LOOP_EVAL_PARAMETERS: rlt::rl::loop::steps::evaluation::Parameters<T, TI, NEXT>{
    static constexpr TI EVALUATION_INTERVAL = 4;
    static constexpr TI NUM_EVALUATION_EPISODES = 10;
    static constexpr TI N_EVALUATIONS = NEXT::CORE_PARAMETERS::STEP_LIMIT / EVALUATION_INTERVAL;
};
using LOOP_CONFIG = rlt::rl::loop::steps::evaluation::Config<LOOP_CORE_CONFIG, LOOP_EVAL_PARAMETERS<LOOP_CORE_CONFIG>>;
using LOOP_STATE = typename LOOP_CONFIG::template State<LOOP_CONFIG>;

DEVICE device;
TI seed = 1337;
LOOP_STATE ls;
rlt::malloc(device, ls);
rlt::init(device, ls, seed);
ls.actor_optimizer.parameters.alpha = 1e-3; // increasing the learning rate leads to faster training of the Pendulum-v1 environment
ls.critic_optimizer.parameters.alpha = 1e-3;

while(!rlt::step(device, ls)){
}

Step: 0/74 Mean return: -1682.39 Mean episode length: 200
Step: 4/74 Mean return: -1419.99 Mean episode length: 200
Step: 8/74 Mean return: -1183.42 Mean episode length: 200
Step: 12/74 Mean return: -1380.93 Mean episode length: 200
Step: 16/74 Mean return: -1277.29 Mean episode length: 200
Step: 20/74 Mean return: -1395.43 Mean episode length: 200
Step: 24/74 Mean return: -1492.18 Mean episode length: 200
Step: 28/74 Mean return: -1057.63 Mean episode length: 200
Step: 32/74 Mean return: -942.946 Mean episode length: 200
Step: 36/74 Mean return: -607.803 Mean episode length: 200
Step: 40/74 Mean return: -530.716 Mean episode length: 200
Step: 44/74 Mean return: -206.616 Mean episode length: 200
Step: 48/74 Mean return: -165.582 Mean episode length: 200
Step: 52/74 Mean return: -160.067 Mean episode length: 200
Step: 56/74 Mean return: -163.664 Mean episode length: 200
Step: 60/74 Mean return: -174.3 Mean episode length: 200
Step: 64/74 Mean return: -220.182 Mean episode length: 200
Step: 68/74 Mean return: -202.481 Mean episode length: 200