Godot Steering AI Framework

This project is a framework to code complex and smooth AI movement in GDScript, using steering behaviors. It works in both 2D and 3D games.

It supports all essential steering behaviors like flee, follow, look at, but also blended behaviors, group behaviors, avoiding neighbors, following a path, following the leader, and much more.

Table of Contents

• Introduction
• The framework
• Getting Started
• Example usage

Introduction

In the 1990s, Craig Reynolds developed algorithms for common AI behaviors. They allowed AI agents to seek out or flee from a target, follow a pre-defined path, or face in a particular direction. They were simple, repeatable tasks that could be broken down into a programming algorithms which made them easy to reuse, maintain, combine and extend.

While an AI agent's next action is based on decision making and planning algorithms, steering behaviors dictate how it will move from one frame to the next. They use available information and calculate where to move at that moment.

Joining these systems together can give complex and graceful movement while also being more efficient than complex path finding algorithms like A*.

The framework

This project is a framework for the Godot game engine. It takes a lot of inspiration from the excellent GDX-AI framework for the LibGDX java-based framework.

Every class in the toolkit is based on Godot's Reference type. There is no need to have a complex scene tree; everything that has to do with the AI's movement can be contained inside movement-oriented classes.

As a short overview, a character is represented by a steering agent; it stores its position, orientation, maximum speeds and current velocity. A steering behavior is associated with a steering agent and calculates a linear and/or angular change in velocity based on its information. The coder then applies that acceleration in whatever ways is appropriate to the character to change its velocity, like RigidBody's apply_impulse, or a KinematicBody's move_and_slide.

Getting Started

The framework's documentation and code reference are both available on the GDQuest website's documents.

More information and resources

• Understanding Steering Behaviors: Breakdowns of various behaviors by Fernando Bevilacqua with graphics and in-depth explanations.
• GDX-AI Wiki: Descriptions of how LibGDX's AI submodule uses steering behaviors with a few graphics. Since this toolkit uses it for inspiration, there will be some similarities.
• RedBlobGames - An excellent resources for complex pathfinding like A*, graph theory, and other algorithms that are game-development related. Steering behaviors are not covered, but for anyone looking to study and bulk up on their algorithms, this is a great place.

Example usage

The fastest way to get started is to look at a sample class that makes use of the toolkit.

The goal of this class is to show how an agent can chase a player and predict where the player will be while also maintaining a distance from them. When the agent’s health is low, it will flee from the player directly. The agent will keep facing the player while it’s chasing them, but will look where it's going while it’s fleeing.

Our game will be in 2D and assumed to be a top-down spaceship game.

You can see the demo in action by running the demos/QuickStartDemo.tscn scene in Godot. There are other demos there that showcase the various behaviors, and the behavior parameters can be tweaked and changed by changing the demo's root node's parameters.

extends KinematicBody2D


# Maximum possible linear velocity
export var speed_max := 450.0
# Maximum change in linear velocity
export var acceleration_max := 50.0
# Maximum rotation velocity represented in degrees
export var angular_speed_max := 240
# Maximum change in rotation velocity represented in degrees
export var angular_acceleration_max := 40

export var health_max := 100
export var flee_health_threshold := 20

var velocity := Vector2.ZERO
var angular_velocity := 0.0
var linear_drag := 0.1
var angular_drag := 0.1

# Holds the linear and angular components calculated by our steering behaviors.
var acceleration := GSAITargetAcceleration.new()


onready var current_health := health_max

# GSAISteeringAgent holds our agent's position, orientation, maximum speed and acceleration.

onready var agent := GSAISteeringAgent.new()


onready var player: Node = get_tree().get_nodes_in_group("Player")[0]
# This assumes that our player class will keep its own agent updated.
onready var player_agent: GSAISteeringAgent = player.agent


# Proximities represent an area with which an agent can identify where neighbors in its relevant
# group are. In our case, the group will feature the player, which will be used to avoid a
# collision with them. We use a radius proximity so the player is only relevant inside 100 pixels
onready var proximity := GSAIRadiusProximity.new(agent, [player_agent], 100).


# GSAIBlend combines behaviors together, calculating all of their acceleration together and adding

# them together, multiplied by a strength. We will have one for fleeing, and one for pursuing,
# toggling them depending on the agent's health. Since we want the agent to rotate AND move, then
# we aim to blend them together.
onready var flee_blend := GSAIBlend.new(agent)

onready var pursue_blend := GSAIBlend.new(agent)


# GSAIPriority will be the main steering behavior we use. It holds sub-behaviors and will pick the  

# first one that returns non-zero acceleration, ignoring any afterwards.
onready var priority := GSAIPriority.new(agent)



func _ready() -> void:
    # ---------- Configuration for our agent ----------
    agent.linear_speed_max = speed_max
    agent.linear_acceleration_max = acceleration_max
    agent.angular_speed_max = deg2rad(angular_speed_max)
    agent.angular_acceleration_max = deg2rad(angular_acceleration_max)
    agent.bounding_radius = calculate_radius($CollisionPolygon2D.polygon)
    update_agent()

    # ---------- Configuration for our behaviors ----------
    # Pursue will happen while the player is in good health. It produces acceleration that takes
    # the agent on an intercept course with the target, predicting its position in the future.
    var pursue := GSAIPursue.new(agent, player_agent)

    pursue.predict_time_max = 1.5

    # Flee will happen while the agent is in bad health, so will start disabled. It produces
    # acceleration that takes the agent directly away from the target with no prediction.
    var flee := GSAIFlee.new(agent, player_agent)


    # AvoidCollision tries to keep the agent from running into any of the neighbors found in its
    # proximity group. In our case, this will be the player if they are close enough.
    var avoid := GSAIAvoidCollisions.new(agent, proximity)


    # Face turns the agent to keep looking towards its target. It will be enabled while the agent
    # is not fleeing due to low health. It tries to arrive 'on alignment' with 0 remaining velocity.
    var face := GSAIFace.new(agent, player_agent)


    # We use deg2rad because the math in the toolkit assumes radians.
    # How close for the agent to be 'aligned', if not exact.
    face.alignment_tolerance = deg2rad(5)
    # When to start slowing down.
    face.deceleration_radius = deg2rad(45)

    # LookWhereYouGo turns the agent to keep looking towards its direction of travel. It will only
    # be enabled while the agent is at low health.
    var look := GSAILookWhereYouGo.new(agent)

    # How close for the agent to be 'aligned', if not exact.
    look.alignment_tolerance = deg2rad(5)
    # When to start slowing down.
    look.deceleration_radius = deg2rad(45)

    # Behaviors that are not enabled produce 0 acceleration.
    # Adding our fleeing behaviors to a blend. The order does not matter.
    flee_blend.is_enabled = false

    flee_blend.add(look, 1)
    flee_blend.add(flee, 1)

    # Adding our pursuit behaviors to a blend. The order does not matter.
    pursue_blend.add(face, 1)
    pursue_blend.add(pursue, 1)

    # Adding our final behaviors to the main priority behavior. The order does matter here.
    # We want to avoid collision with the player first, flee from the player second when enabled,
    # and pursue the player last when enabled.
    priority.add(avoid)
    priority.add(flee_blend)
    priority.add(pursue_blend)


func _physics_process(delta: float) -> void:
    # Make sure any change in position and speed has been recorded.
    update_agent()

    if current_health <= flee_health_threshold:
        pursue_blend.is_enabled = false

        flee_blend.is_enabled = true


    # Calculate the desired acceleration.
    priority.calculate_steering(acceleration)

    # We add the discovered acceleration to our linear velocity. The toolkit does not limit
    # velocity, just acceleration, so we clamp the result ourselves here.
    velocity = (velocity + Vector2(
                acceleration.linear.x, acceleration.linear.y)
    ).clamped(agent.linear_speed_max)

    # This applies drag on the agent's motion, helping it to slow down naturally.
    velocity = velocity.linear_interpolate(Vector2.ZERO, linear_drag)

    # And since we're using a KinematicBody2D, we use Godot's excellent move_and_slide to actually
    # apply the final movement, and record any change in velocity the physics engine discovered.
    velocity = move_and_slide(velocity)

    # We then do something similar to apply our agent's rotational speed.
    angular_velocity = clamp(
            angular_velocity + acceleration.angular,
            -agent.angular_speed_max,
            agent.angular_speed_max
    )
    # This applies drag on the agent's rotation, helping it slow down naturally.
    angular_velocity = lerp(angular_velocity, 0, angular_drag)
    rotation += angular_velocity * delta


# In order to support both 2D and 3D, the toolkit uses Vector3, so the conversion is required
# when using 2D nodes. The Z component can be left to 0 safely.
func update_agent() -> void:
    agent.position.x = global_position.x
    agent.position.y = global_position.y
    agent.orientation = rotation
    agent.linear_velocity.x = velocity.x
    agent.linear_velocity.y = velocity.y
    agent.angular_velocity = angular_velocity


# We calculate the radius from the collision shape - this will approximate the agent's size in the
# game world, to avoid collisions with the player.
func calculate_radius(polygon: PoolVector2Array) -> float:
    var furthest_point := Vector2(-INF, -INF)
    for p in polygon:
        if abs(p.x) > furthest_point.x:
            furthest_point.x = p.x
        if abs(p.y) > furthest_point.y:
            furthest_point.y = p.y
    return furthest_point.length()


func damage(amount: int) -> void:
    current_health -= amount
    if current_health <= 0:
        queue_free()