🏋️‍♀️ bevy_rl

Reinforcement Learning for Bevy Engine

🏗️ Build 🤔 Reinforcement Learning 🏋🏿‍♂️ Gym environments with 🕊 Bevy engine to train 👾 AI agents that 💡 can learn from 📺 screen pixels or defined obeservation state.

Compatibility

📝Features

• Set of APIs to implement OpenAI Gym interface, such as reset, step, render, close and associated simulator states
• Multi-Agent support
• Rendering screen pixels to RAM buffer — for training agents with raw pixels
• REST API to control agents

👩‍💻 Usage

1. Define Action and Obeservation Space

Observation space needs to be Serializable for REST API to work.

// Action space
#[derive(Default)]
pub struct Actions {
    // actuator_signals: [f32; 3],
}

// Observation space
#[derive(Default, Serialize, Clone)]
pub struct EnvironmentState {
    // agent_coords: [(f32, f32); 16],
}

2. Enable AI Gym Plugin

Width and height should exceed 256, otherwise wgpu will panic.

// Setup bevy_rl
let ai_gym_state = AIGymState::<Actions, State>::new(AIGymSettings {
    width: u32,              // Width and height of the screen
    height: u32,             // ...
    num_agents: 1,           // Number of agents — each will get a camera handle
    render_to_buffer: false, // You can disable rendering to buffer
    pause_interval: 0.01,    // 100 Hz
    ..default()
});
app.insert_resource(ai_gym_state)
    .add_plugin(AIGymPlugin::<Actions, EnvironmentState>::default());

2.1 (Optional) Enable Rendering to Buffer

If your environment wants to export raw pixels, you will need to attach a render target to each camera you want to export them from. Render targets are copied each frame from GPU memory to RAM buffers so that they can be accessed with REST API.

pub(crate) fn spawn_cameras(
    ai_gym_state: Res<AIGymState<Actions, EnvironmentState>>,
) {
    let mut ai_gym_state = ai_gym_state.lock().unwrap();
    let ai_gym_settings = ai_gym_state.settings.clone();

    for i in 0..ai_gym_settings.num_agents {
        let render_image_handle = ai_gym_state.render_image_handles[i as usize].clone();
        let render_target = RenderTarget::Image(render_image_handle);
        let camera_bundle = Camera3dBundle {
            camera: Camera {
                target: render_target,  // Render target is baked in bevy_rl and used to export pixels
                priority: -1,           // set to -1 to render at the firstmost pass
                ..default()
            },
            ..default()
        };
        commands.spawn(camera_bundle);
    }
}

4. Handle bevy_rl events

bevy_rl will communicate with your environment through events. You can use EventReader to read events and respond to them. Those event are from REST API or from a timer that pauses the simulation with given interval (AIGymSettings.pause_interval).

Here's example of how to handle those events:

// EventPauseResume
fn bevy_rl_pause_request(
    mut pause_event_reader: EventReader<EventPauseResume>,
    ai_gym_state: Res<AIGymState<Actions, State>>,
) {
    for _ in pause_event_reader.iter() {
        // Pause simulation (physics engine)
        // ...
        // Collect state into serializable struct
        let env_state = EnvironmentState(...);
        // Set bevy_rl gym state
        let mut ai_gym_state = ai_gym_state.lock().unwrap();
        ai_gym_state.set_env_state(env_state);
    }
}

// EventControl
fn bevy_rl_control_request(
    mut pause_event_reader: EventReader<EventControl>,
    mut simulation_state: ResMut<State<SimulationState>>,
) {
    for control in pause_event_reader.iter() {
        let unparsed_actions = &control.0;
        for i in 0..unparsed_actions.len() {
            if let Some(unparsed_action) = unparsed_actions[i].clone() {
                let action: Vec<f64> = serde_json::from_str(&unparsed_action).unwrap();
                // Pass control inputs to your agents
                // ...
            }
        }
        // Resume simulation (physics engine)
        // ...
        // Return to running state; note that it uses pop/push to avoid
        // entering `SystemSet::on_enter(SimulationState::Running)` which initialized game world anew
        simulation_state.pop().unwrap();
    }
}

/// Handle bevy_rl::EventReset
pub(crate) fn bevy_rl_reset_request(
    mut reset_event_reader: EventReader<EventReset>,
    mut commands: Commands,
    mut walls: Query<Entity, &Wall>,
    mut players: Query<(Entity, &Actor)>,
    mut simulation_state: ResMut<State<SimulationState>>,
    ai_gym_state: Res<AIGymState<Actions, EnvironmentState>>,
) {
    if reset_event_reader.iter().count() == 0 {
        return;
    }

    // Reset envrionment state here

    // Return simulation in Running state
    simulation_state.set(SimulationState::Running).unwrap();

    // tell bevy_rl that environment is reset and return response to REST API
    let ai_gym_state = ai_gym_state.lock().unwrap();
    ai_gym_state.send_reset_result(true);
}

Register systems to handle bevy_rl events.

app.add_system_set(
    SystemSet::on_update(SimulationState::PausedForControl)
        .with_system(bevy_rl_control_request)
        .with_system(bevy_rl_reset_request)
        .with_system(bevy_rl_pause_request),
);

💻 AIGymState API

Those methods are available on AIGymState resource. You should use them to alter bevy_rl internal state.

🌐 REST API

Accessing bevy_rl-enabled environment is possible through REST API. You can use any HTTP client to communicate with it. Here's a list of available endpoints:

One would wrap those endpoints into a client library (python) to make it easier to use.

✍️ Examples

• bevy_rl_shooter — example FPS project
• bevy_quadruped_neural_control — quadruped locomotion with bevy_mujoco and bevy_rl