Controllers

A controller is a higher-level tool computing the motion of a body from what your application asks for, rather than leaving it entirely to the forces of the simulation. Rapier provides the following controllers, each being an ordinary Python object owned by your application (it doesn’t belong to the PhysicsWorld, and is given the parts of the world it needs when it is used):

• The character controller, moving a kinematic body along the obstacles of the scene.
• The vehicle controller, driving a car-like body on ray-casted wheels.
• The PID controller, steering a dynamic body toward a target pose at the velocity level.

Character controller

Most games involve bodies behaving in ways that defy the laws of physics: floating platforms, elevators, playable characters, etc. This is why kinematic bodies exist: they offer a total control over the body’s trajectory since they are completely immune to forces or impulses (like gravity, contacts, joints).

But this control comes at a price: it is up to the user to take any obstacle into account by running custom collision-detection operations manually and update the trajectory accordingly. This can be very difficult. Detecting obstacles usually rely on ray-casting or shape-casting, used to adjust the trajectory based on the potential contact normals. Often, multiple ray or shape-casts are needed, and the trajectory adjustment code isn’t straightforward.

The Kinematic Character Controller (which we will abbreviate to character controller) is a higher-level tool that will emit the proper ray-casts and shape-casts to adjust the user-defined trajectory based on obstacles. The well-known move-and-slide operation is the main feature of a character controller.

note

Despite its name, a character controller can also be used for moving objects that are not characters. For example, a character controller may be used to move a platform. In the rest of this guide, we will use the word character to designate whatever you would like to move using the character controller.

Rapier provides a built-in general-purpose character controller implementation. It allows you to easily:

• Stop at obstacles.
• Slide on slopes that are not to steep.
• Climb stairs automatically.
• Walk over small obstacles.
• Interact with moving platforms.

Despite the fact that this built-in character controller is designed to be generic enough to serve as a good starting point for many common use-cases, character-control (especially for the player’s character itself) is often very game-specific. Therefore the builtin character controller may not work perfectly out-of-the-box for all game types.

Setup and usage

The character controller implementation is exposed as the KinematicCharacterController class. This class only contains information about the character controller’s behavior. It does not contain any collider-specific or rigid-body-specific information like handles, velocities, positions, etc. Therefore, the same instance of KinematicCharacterController can be used to control multiple rigid-bodies/colliders if they rely on the same set of parameters. Its settings (detailed in the next sections) are properties that can also be given as keyword arguments to its constructor (e.g. KinematicCharacterController(slide=False, max_slope_climb_angle=0.5)). The KinematicCharacterController exposes only two methods:

• move_shape is responsible for calculating the possible movement of a character based on the desired movement, obstacles, and character controller options. It is given the timestep length, the rigid-body set and the collider set (both unused and kept for backward compatibility: they can be None, or must be the sets of the query pipeline), the query pipeline of the world containing the obstacles (e.g. world.query_pipeline), the shape of the character (a SharedShape, e.g. the shape of its Collider), its current pose, the desired translation, and optionally the QueryFilter selecting the obstacles (see the filtering section) and a callback receiving the collisions (see the collisions section).
• solve_character_collision_impulses is detailed in the collisions section.

# The translation we would like to apply if there were no obstacles.
desired_translation = (1.0, -2.0, 3.0)
# Create the character controller, here with the default configuration.
character_controller = rp.KinematicCharacterController()
# Make sure the character we are trying to move isn’t considered an obstacle.
query_filter = rp.QueryFilter().exclude_rigid_body(rigid_body_handle)
# Calculate the possible movement.
corrected_movement = character_controller.move_shape(
    dt,  # The timestep length (can be set to world.integration_parameters.dt).
    None,  # The rigid-body set, unused: the one of the query pipeline is used.
    None,  # The collider set, unused: the one of the query pipeline is used.
    world.query_pipeline,  # The query pipeline containing the obstacles.
    character_shape,  # The character’s shape.
    character_pos,  # The character’s initial position.
    desired_translation,
    query_filter,  # The obstacles to consider.
)
# TODO: apply the `corrected_movement.translation` to the rigid-body or collider based on the rules described below.

The obstacles are taken at the positions they had at the end of the last PhysicsWorld.step (or the last PhysicsWorld.update_query_pipeline). The returned EffectiveCharacterMovement contains the corrected movement (its translation property), as well as whether the character touches the ground at its final position (its grounded property), and whether it is sliding down a slope that is too steep to climb (its is_sliding_down_slope property). The corrected movement isn’t applied automatically: the recommended way to update the character’s position depends on its representation:

• A collider not attached to any rigid-body: set the collider’s position directly (with its translation property) to the corrected movement added to its current position.
• A velocity-based kinematic rigid-body: set its velocity (with its linvel property) to the computed movement divided by the timestep length.
• A position-based kinematic rigid-body: set its next kinematic position (with RigidBody.set_next_kinematic_translation) to the corrected movement added to its current position.

info

The character’s shape may be any shape supported by Rapier. However, it is recommended to either use a cuboid, a ball, or a capsule since they involve less computations and less numerical approximations.

warning

The built-in character controller does not support rotational movement. It only supports translations.

Character offset

For performance and numerical stability reasons, the character controller will attempt to preserve a small gap between the character shape and the environment. This small gap is named offset and acts as a small margin around the character shape. A good value for this offset is something sufficiently small to make the gap unnoticeable, but sufficiently large to avoid numerical issues (if the character seems to get stuck inexplicably, try increasing the offset).

The offset is set with the offset property. Like every length of the character controller, it is given as a CharacterLength: either an absolute length (created by CharacterLength.absolute), or a factor multiplied by the size of the character’s shape (created by CharacterLength.relative). A plain float is also accepted, and interpreted as an absolute length.

# The character offset is set to 0.01.
character_controller.offset = rp.CharacterLength.absolute(0.01)
# The character offset is set to 0.01 multiplied by the shape’s height.
character_controller.offset = rp.CharacterLength.relative(0.01)

warning

It is not recommended to change the offset after the creation of the character controller.

Up vector

The up vector instructs the character controller of what direction should be considered vertical. The horizontal plane is the plane orthogonal to this up vector. There are two equivalent ways to evaluate the slope of the floor: by taking the angle between the floor and the horizontal plane (in 2D), or by taking the angle between the up-vector and the normal of the floor (in 2D and 3D). By default, the up vector is the positive y axis, but it can be modified to be any (unit) vector that suits the application.

The up vector is set with the up property (it isn’t normalized automatically, so it must be given as a unit vector):

# Set the up-vector to the positive X axis.
character_controller.up = (1.0, 0.0, 0.0)

Slopes

If sliding is enabled, the character can automatically climb slopes if they are not too steep, or slide down slopes if they are too steep. Sliding is configured by the following parameters:

• The max slope climb angle: if the angle between the slope to climb and the horizontal floor is larger than this value, then the character won’t be able to slide up this slope.
• The min slope slide angle: if the angle between the slope and the horizontal floor is smaller than this value, then the vertical component of the character’s movement won’t result in any sliding.

info

As always in Rapier, angles are specified in radians.

The angles are set with the max_slope_climb_angle and min_slope_slide_angle properties. Sliding itself is enabled by default, and can be disabled by setting the slide property to False.

# Don’t allow climbing slopes larger than 45 degrees.
character_controller.max_slope_climb_angle = math.radians(45.0)
# Automatically slide down on slopes smaller than 30 degrees.
character_controller.min_slope_slide_angle = math.radians(30.0)

Stairs and small obstacles

If enabled, the autostep setting allows the character to climb stairs automatically and walk over small obstacles. Autostepping requires the following parameters:

• The maximum height the character can step over. If the vertical movement needed to step over this obstacle is larger than this value, then the character will be stopped by the obstacle.
• The minimum (horizontal) width available on top of the obstacle. If, after the character is teleported on top of the obstacle, it cannot move forward by a distance larger than this minimum width, then the character will just be stopped by the obstacle (without being moved to the top of the obstacle).
• Whether or not autostepping is enabled for dynamic bodies. If it is not enabled for dynamic bodies, the character won’t attempt to automatically step over small dynamic bodies. Disabling this can be useful if we want the character to push these small objects (see collisions) instead of just stepping over them.

The following depicts (top) one configuration where all the autostepping conditions are satisfied, and, (bottom) two configurations where these conditions are not all satisfied (left: because the width of the step is too small, right: because the height of the step is too large):

info

Autostepping will only activate if the character is touching the floor right before the obstacle. This prevents the player from being teleported on to of a platform while it is in the air.

Autostepping is configured with the autostep property: either None (the default) to disable it, or a CharacterAutostep created from the maximum height and minimum width (both as CharacterLength, see the character offset), and whether it is enabled for dynamic bodies:

# Set autostep to None to disable it.
character_controller.autostep = None
# Autostep if the step height is smaller than 0.5, and its width larger than 0.2.
character_controller.autostep = rp.CharacterAutostep(
    max_height=rp.CharacterLength.absolute(0.5),
    min_width=rp.CharacterLength.absolute(0.2),
    include_dynamic_bodies=True,
)
# Autostep if the step height is smaller than 0.3 multiplied by the character’s height,
# and its width larger than 0.5 multiplied by the character’s width (i.e. half the character’s
# width).
character_controller.autostep = rp.CharacterAutostep(
    max_height=rp.CharacterLength.relative(0.3),
    min_width=rp.CharacterLength.relative(0.5),
    include_dynamic_bodies=True,
)

Snap-to-ground

If enabled, snap-to-ground will force the character to stick to the ground if the following conditions are met simultaneously:

• At the start of the movement, the character touches the ground.
• The movement has a slight downward component.
• At the end of the desired movement, the character would be separated from the ground by a distance smaller than the distance provided by the snap-to-ground parameter.

If these conditions are met, the character is automatically teleported down to the ground at the end of its motion. Typical usages of snap-to-ground include going downstairs or remaining in contact with the floor when moving downhill.

Snap-to-ground is configured with the snap_to_ground property: either None to disable it, or the snapping distance as a CharacterLength (see the character offset). It is enabled by default, with a distance of CharacterLength.relative(0.2):

# Set snap-to-ground to None to disable it.
character_controller.snap_to_ground = None
# Snap to the ground if the vertical distance to the ground is smaller than 0.5.
character_controller.snap_to_ground = rp.CharacterLength.absolute(0.5)
# Snap to the ground if the vertical distance to the ground is smaller than 0.2 times the character’s height.
character_controller.snap_to_ground = rp.CharacterLength.relative(0.2)

Filtering

It is possible to let the character controller ignore some obstacles. This is achieved by configuring the filter argument of the KinematicCharacterController.move_shape method. This QueryFilter class is detailed in the scene query filters section:

• Its flags (a QueryFilterFlags, also set by constructors like QueryFilter.exclude_dynamic()) allow you to exclude whole families of obstacles (e.g. all the colliders attached to dynamic rigid-bodies).
• Its groups method allows you to filter based on the colliders collision groups.
• Its exclude_collider and exclude_rigid_body methods exclude one collider, and all the colliders attached to one rigid-body.
• Its predicate method sets an optional callable to filter-out colliders based on user-defined rules. It is given the ColliderHandle and the Collider of each potential obstacle, and returns True to keep it as an obstacle.

warning

If the character-controller is used to move a collider (and the rigid-body it may be attached to) that is present in the physics scene, the filters must be used to exclude that collider (and that rigid-body) from the set of obstacles (with QueryFilter.exclude_collider and QueryFilter.exclude_rigid_body) to prevent the character from colliding with itself.

Collisions

As the character moves along its path, it will hit grounds and obstacles before sliding or stepping on them. Knowing what collider was hit on this path, and where the hit took place, can be valuable to apply various logic (custom forces, sound effects, etc.) This is why a set of character collision events are collected during the calculation of its trajectory.

info

The character collision events are given in chronological order. For example, if, during the resolution of the character motion, the character hits an obstacle A, then slides against it, and then hits another obstacle B. The collision with A will be reported first, and the collision with B will be reported second.

The character collisions are given to the optional events_callback argument of KinematicCharacterController.move_shape: a callable invoked with each CharacterCollision while the movement is being calculated (an exception raised by that callable is re-raised by move_shape):

character_controller = rp.KinematicCharacterController()

def on_collision(collision):
    # Handle or collect the collision in this callback.
    pass

# Give a callback to handle or collect the collisions while
# the character is being moved.
character_controller.move_shape(
    dt,
    None,
    None,
    world.query_pipeline,
    character_shape,
    character_pos,
    desired_translation,
    query_filter,
    events_callback=on_collision,
)

Each collision gives the handle of the collider hit, the pose of the character at the time of the hit (character_pos), as well as the parts of the desired translation already applied (translation_applied) and remaining (translation_remaining) at that time. Its hit property is a ShapeCastHit, i.e., it has the same form as the result of a shape-casting: its first witness point and normal are on the obstacle, in world-space, whereas its second witness point and normal are on the character, in its local-space.

Unless dynamic bodies are filtered-out by the character controller’s filters, they may be hit during the resolution of the character movement. If that happens, these dynamic bodies will generally not react to (i.e. not be pushed by) the character because the character controller’s offset prevents actual contacts from happening.

In these situations forces need to be applied manually to this rigid-bodies. The character controller can apply these forces for you if needed:

# First, collect all the collisions.
collisions = []
character_controller.move_shape(
    dt,
    None,
    None,
    world.query_pipeline,
    character_shape,
    character_pos,
    desired_translation,
    query_filter,
    events_callback=collisions.append,
)
# Then, let the character controller solve (and apply) the collision impulses
# to the dynamic rigid-bodies hit along its path.
character_controller.solve_character_collision_impulses(
    dt,
    world.rigid_bodies,
    world.colliders,
    world.query_pipeline,
    character_shape,
    None,  # Unused: each collision stores the pose of the character when it happened.
    character_mass,
    collisions,
    query_filter,
)

The impulses are computed from the collisions collected during the move_shape call (each of them stores the pose of the character when it happened, so the character_pos argument is unused and can be None), and applied to the rigid-bodies of the given rigid-body set (which, like the collider set, must be the one of the query pipeline). The shape and filter given to solve_character_collision_impulses should be the same as the ones given to move_shape. Unlike move_shape, this calls the predicate of the filter, if any, once for every collider before applying the impulses, while the sets are locked: it must not modify them. The mass of the character is given explicitly: it can be, for example, the mass of the rigid-body it is attached to (as given by its RigidBody.mass property).

Gravity

Since you are responsible for providing the movement vector to the character controller at each frame, it is up to you to emulate gravity by adding a downward component to that movement vector.

Vehicle controller

Simulating a car with rigid-bodies and joints, for example using one rigid-body per wheel attached to the chassis by a joint, is possible but can be difficult to control for games requiring non-realistic vehicles. This is why Rapier provides a vehicle controller (the current implementation was ported from the btRaycastVehicle of Bullet). The vehicle is a single rigid-body modeling its chassis, and its wheels are only represented by ray-casts pushing that body along a spring-like suspension.

Setup

The chassis is created like any other dynamic rigid-body, and the wheels are added to the controller afterwards. Each wheel is given its position on the chassis, the direction of its suspension (the direction of its ray-cast), its axle, the rest length of its suspension, and its radius. The WheelTuning shared by the wheels is what makes the vehicle feel heavy or light by controlling the elastic properties (stiffness and damping) of the suspension, as well as the grip of the wheels:

The vehicle controller is the DynamicRayCastVehicleController class, created from the handle of the chassis' dynamic rigid-body. The wheels are added with its add_wheel method, which returns the index of the new wheel (starting from zero, in insertion order). The WheelTuning given to each wheel is created with keyword arguments overriding its default values (as given by WheelTuning.default()):

# The chassis is an ordinary dynamic rigid-body.
hw = 0.3
hh = 0.15
chassis_handle = world.add_body(
    rp.RigidBody.dynamic(translation=(0.0, 1.0, 0.0)),
    colliders=[rp.Collider.cuboid(hw * 2.0, hh, hw).density(100.0)],
)

# The tuning shared by the wheels: the suspension and the grip. The
# parameters that aren't given keep their default values.
tuning = rp.WheelTuning(suspension_stiffness=100.0, suspension_damping=10.0)

vehicle = rp.DynamicRayCastVehicleController(chassis_handle)
wheel_positions = [
    (hw * 1.5, -hh, hw),
    (hw * 1.5, -hh, -hw),
    (-hw * 1.5, -hh, hw),
    (-hw * 1.5, -hh, -hw),
]

for position in wheel_positions:
    # The position of the wheel, the direction its suspension pushes along, its axle, the
    # rest length of its suspension, and its radius; all in the local frame of the chassis.
    vehicle.add_wheel(position, (0.0, -1.0, 0.0), (0.0, 0.0, 1.0), hh, hh / 4.0, tuning)

By default, the vehicle moves forward along the local x axis of the chassis, and its local y axis points upward. Other axes can be selected with the index_forward_axis and index_up_axis properties of the controller (0, 1, and 2 standing for the x, y, and z axes).

Driving the vehicle

A vehicle is driven by giving each of its wheels an engine force, a brake force, and a steering angle. The controller is then updated before each timestep, which is when the ray-casts are made and when the resulting suspension and friction forces are applied to the chassis. It is strongly recommended to exclude the chassis from the ray-casts, otherwise the ray might hit it and be misinterpreted as being the floor.

The engine force, brake, and steering angle of a wheel are set with the apply_engine_force, set_brake, and set_steering methods of the controller (from the index of the wheel), and the controller is updated with its update_vehicle method before each PhysicsWorld.step. The colliders attached to the chassis are always excluded from the ray-casts, and the other obstacles can be filtered with the optional QueryFilter given to update_vehicle (see the scene query filters). The predicate of that filter, if any, is called once for every collider before the update, while the sets are locked: it must not modify them.

for _ in range(200):
    # The vehicle is driven by setting the engine force, the brake, and the steering angle of
    # its wheels (from their indices). Here the two front wheels are the driving and steering ones.
    vehicle.apply_engine_force(0, 30.0)
    vehicle.set_steering(0, 0.2)
    vehicle.apply_engine_force(1, 30.0)
    vehicle.set_steering(1, 0.2)

    # The wheels are ray-casted against the scene: the chassis itself is always excluded from
    # these ray-casts, and every other dynamic body is generally excluded too.
    vehicle.update_vehicle(
        world.integration_parameters.dt,
        world.rigid_bodies,
        world.colliders,
        world.query_pipeline,
        rp.QueryFilter.exclude_dynamic(),
    )

    world.step()

print("Vehicle speed:", vehicle.current_vehicle_speed)

note

The state of each wheel after an update (whether it touches the floor, the compression of its suspension, its rotation angle, etc.) is readable from the controller. This is what the rendering of the wheels can be based on since there is no actual per-wheel rigid-bodies to read their state from.

The state of the wheels is given by the wheels method of the controller (or by its wheel method for a single wheel) as a list of Wheel objects. These are copies: modifying them doesn’t affect the vehicle. Each wheel gives its world-space center, the world-space directions of its suspension and axle, its rotation angle, the force of its suspension (wheel_suspension_force), and its raycast_info (a RayCastInfo): the length of its suspension (suspension_length) and its contact with the ground (is_in_contact, the collider touched ground_object, and the world-space contact point contact_point_ws and normal contact_normal_ws). The speed of the vehicle along its forward axis is given by the current_vehicle_speed property of the controller (in meters per second, or with the current_speed_km_hour method in kilometers per hour).

# The wheels are copies of the state of the wheels after the last update.
for wheel in vehicle.wheels():
    contact = wheel.raycast_info
    print(
        "Wheel center:", wheel.center,  # World-space center of the wheel.
        "axle:", wheel.axle,  # World-space direction of its axle.
        "rotation:", wheel.rotation,  # Rotation angle around its axle.
        "suspension length:", contact.suspension_length,
        "touches the ground:", contact.is_in_contact,
        "ground collider:", contact.ground_object,  # None if it doesn’t touch the ground.
    )

PID controller

It is generally not recommended to move a rigid-body by setting its pose directly: teleporting it would ignore every obstacle on the way. The recommended alternative is generally to push it with a force (or an impulse) that is strong enough to reach the target. However, pushing with a single constant force/impulse will generally overshoot the target. Thus, ideally, the force or impulse should be carefully selected and updated each frame as the rigid-body gets closer to its target.

This is what a PID controller (Proportional-Integral-Derivative) is designed to calculate: given the target pose, it computes the ideal velocity change bringing the body closer to it. This is the building block of the velocity-based character controllers, but it is useful for anything that must follow a target without being teleported: a dynamic moving platform, an object held by the player, a following camera, etc.

info

The gains of the controller are what makes it reach its target quickly or smoothly. The proportional gain is applied to the position errors and is usually set to a multiple of the inverse of the timestep length (e.g. $60$ for a timestep of $1 / 60$ seconds). The derivative gain is applied to the velocity errors and is usually set in $[0, 1]$, where $0$ means no damping and $1$ means that the velocity errors are corrected within a single timestep.

The PID controller is the PidController class. It is created with a proportional gain of $60$, an integral gain of $1$, and a derivative gain of $0.8$, on every coordinate axis (all of them being controlled), unless other gains are given to the Kp, Ki, and Kd arguments of its constructor (either a single float for every linear and angular coordinate axis, or one gain per coordinate axis). They can be read and modified afterwards, per coordinate axis, with its lin_kp, ang_kp, lin_ki, ang_ki, lin_kd, and ang_kd properties (each one is a Vec3, and can be set from a vector or from a single float for every axis). At each frame, PidController.rigid_body_correction computes the velocity correction (a PidCorrection, with its linear and angular parts) bringing a rigid-body closer to its target pose (and to its target velocities, given as an optional RigidBodyVelocity to its target_vels argument, zero by default): it is up to you to add it to the linvel and angvel of that rigid-body before the next PhysicsWorld.step.

The coordinate axes (linear and/or angular) controlled by the controller can be selected in order, for example, to only control the translations of a body while leaving its rotations to the simulation (with the axes argument of its constructor or its axes property, a combination of the AxesMask flags):

# The proportional, integral, and derivative gains of the controller, acting on the linear
# axes only: the body is pushed toward its target without its rotation being controlled.
axes = rp.AxesMask.LIN_X | rp.AxesMask.LIN_Y | rp.AxesMask.LIN_Z
pid = rp.PidController(axes=axes, Kp=60.0, Ki=0.0, Kd=0.8)
target = rp.Isometry3.from_translation(3.0, 2.0, 0.0)

for _ in range(200):
    dt = world.integration_parameters.dt
    body = world.rigid_bodies[body_handle]
    # The correction is the velocity change bringing the body closer to its target pose
    # (and to its target velocities, zero here).
    correction = pid.rigid_body_correction(dt, body, target, target_vels=rp.RigidBodyVelocity())
    body.linvel = body.linvel + correction.linear
    body.angvel = body.angvel + correction.angular

    world.step()

note

The integral part of the controller accumulates the position errors of the previous timesteps, which is what allows it to compensate a permanent perturbation (e.g. the gravity applied to a hovering body). These accumulated errors are readable with the PidController.lin_integral and PidController.ang_integral properties. This is also what makes PidController.rigid_body_correction modify the controller, and what has to be reset with PidController.reset whenever the controller is given a target it never had a chance to reach. The PdController is the variant without that integral part (its constructor takes no Ki argument): its rigid_body_correction method doesn’t modify it (and doesn’t need the timestep length), and its behavior is generally good enough for games.