Merge branch 'james/rrt-pathfinding' into 'master'

Initial RRT flight pathfinding

See merge request veloren/veloren!2773
This commit is contained in:
Joshua Barretto 2021-09-19 19:30:31 +00:00
commit e098a38d8e
3 changed files with 459 additions and 19 deletions

1
Cargo.lock generated
View File

@ -5873,6 +5873,7 @@ dependencies = [
"fxhash",
"hashbrown 0.11.2",
"indexmap",
"kiddo",
"lazy_static",
"num-derive",
"num-traits",

View File

@ -11,6 +11,7 @@ simd = ["vek/platform_intrinsics"]
bin_csv = ["ron", "csv", "structopt"]
bin_graphviz = ["petgraph"]
bin_cmd_doc_gen = []
rrt_pathfinding = ["kiddo"]
default = ["simd"]
@ -63,6 +64,8 @@ csv = { version = "1.1.3", optional = true }
structopt = { version = "0.3.13", optional = true }
# graphviz exporters
petgraph = { version = "0.5.1", optional = true }
# K-d trees used for RRT pathfinding
kiddo = { version = "0.1", optional = true }
# Data structures
hashbrown = { version = "0.11", features = ["rayon", "serde", "nightly"] }
@ -103,4 +106,4 @@ required-features = ["bin_graphviz"]
[[bin]]
name = "cmd_doc_gen"
required-features = ["bin_cmd_doc_gen"]
required-features = ["bin_cmd_doc_gen"]

View File

@ -5,7 +5,13 @@ use crate::{
};
use common_base::span;
use hashbrown::hash_map::DefaultHashBuilder;
use rand::prelude::*;
#[cfg(rrt_pathfinding)] use hashbrown::HashMap;
#[cfg(rrt_pathfinding)]
use kiddo::{distance::squared_euclidean, KdTree}; // For RRT paths (disabled for now)
#[cfg(rrt_pathfinding)]
use rand::distributions::Uniform;
use rand::{thread_rng, Rng};
#[cfg(rrt_pathfinding)] use std::f32::consts::PI;
use std::iter::FromIterator;
use vek::*;
@ -135,9 +141,9 @@ impl Route {
})
});
// Map position of node to middle of block
let next_tgt = next0.map(|e| e as f32) + Vec3::new(0.5, 0.5, 0.0);
let closest_tgt = next_tgt.map2(pos, |tgt, pos| pos.clamped(tgt.floor(), tgt.ceil()));
// Determine whether we're close enough to the next to to consider it completed
let dist_sqrd = pos.xy().distance_squared(closest_tgt.xy());
if dist_sqrd
@ -312,9 +318,9 @@ impl Route {
Some((
tgt - pos,
// Control the entity's speed to hopefully stop us falling off walls on sharp corners.
// This code is very imperfect: it does its best but it can still fail for particularly
// fast entities.
// Control the entity's speed to hopefully stop us falling off walls on sharp
// corners. This code is very imperfect: it does its best but it
// can still fail for particularly fast entities.
straight_factor * traversal_cfg.slow_factor + (1.0 - traversal_cfg.slow_factor),
))
.filter(|(bearing, _)| bearing.z < 2.1)
@ -336,6 +342,9 @@ pub struct Chaser {
}
impl Chaser {
/// Returns bearing and speed
/// Bearing is a Vec3<f32> dictating the direction of movement
/// Speed is an f32 between 0.0 and 1.0
pub fn chase<V>(
&mut self,
vol: &V,
@ -386,12 +395,17 @@ impl Chaser {
.and_then(|(r, _)| r.traverse(vol, pos, vel, &traversal_cfg))
}
} else {
// There is no route found yet
None
};
// If a bearing has already been determined, use that
if let Some((bearing, speed)) = bearing {
Some((bearing, speed))
} else {
// Since no bearing has been determined yet, a new route will be
// calculated if the target has moved, pathfinding is not complete,
// or there is no route
let tgt_dir = (tgt - pos).xy().try_normalized().unwrap_or_default();
// Only search for a path if the target has moved from their last position. We
@ -406,7 +420,12 @@ impl Chaser {
{
self.last_search_tgt = Some(tgt);
let (path, complete) = find_path(&mut self.astar, vol, pos, tgt, &traversal_cfg);
// NOTE: Enable air paths when air braking has been figured out
let (path, complete) = /*if cfg!(rrt_pathfinding) && traversal_cfg.can_fly {
find_air_path(vol, pos, tgt, &traversal_cfg)
} else */{
find_path(&mut self.astar, vol, pos, tgt, &traversal_cfg)
};
self.route = path.map(|path| {
let start_index = path
@ -429,19 +448,40 @@ impl Chaser {
)
});
}
let walking_towards_edge = (-3..2).all(|z| {
vol.get(
(pos + Vec3::<f32>::from(tgt_dir) * 2.5).map(|e| e as i32) + Vec3::unit_z() * z,
)
.map(|b| b.is_air())
.unwrap_or(false)
});
if !walking_towards_edge || traversal_cfg.can_fly {
Some(((tgt - pos) * Vec3::new(1.0, 1.0, 0.0), 1.0))
// Start traversing the new route if it exists
if let Some(bearing) = self
.route
.as_mut()
.and_then(|(r, _)| r.traverse(vol, pos, vel, &traversal_cfg))
{
Some(bearing)
} else {
None
// At this point no route is available and no bearing
// has been determined, so we start sampling terrain.
// Check for falling off walls and try moving straight
// towards the target if falling is not a danger
let walking_towards_edge = (-3..2).all(|z| {
vol.get(
(pos + Vec3::<f32>::from(tgt_dir) * 2.5).map(|e| e as i32)
+ Vec3::unit_z() * z,
)
.map(|b| b.is_air())
.unwrap_or(false)
});
// Enable when airbraking/flight is figured out
/*if traversal_cfg.can_fly {
Some(((tgt - pos) , 1.0))
} else */
if !walking_towards_edge || traversal_cfg.can_fly {
Some(((tgt - pos) * Vec3::new(1.0, 1.0, 0.0), 1.0))
} else {
// This is unfortunately where an NPC will stare blankly
// into space. No route has been found and no temporary
// bearing would suffice. Hopefully a route will be found
// in the coming ticks.
None
}
}
}
}
@ -631,3 +671,399 @@ where
PathResult::Pending => (None, false),
}
}
// Enable when airbraking/sensible flight is a thing
#[cfg(rrt_pathfinding)]
fn find_air_path<V>(
vol: &V,
startf: Vec3<f32>,
endf: Vec3<f32>,
traversal_cfg: &TraversalConfig,
) -> (Option<Path<Vec3<i32>>>, bool)
where
V: BaseVol<Vox = Block> + ReadVol,
{
let radius = traversal_cfg.node_tolerance;
let mut connect = false;
let total_dist_sqrd = startf.distance_squared(endf);
// First check if a straight line path works
if vol
.ray(startf + Vec3::unit_z(), endf + Vec3::unit_z())
.until(Block::is_opaque)
.cast()
.0
.powi(2)
>= total_dist_sqrd
{
let mut path = Vec::new();
path.push(endf.map(|e| e.floor() as i32));
connect = true;
(Some(path.into_iter().collect()), connect)
// Else use RRTs
} else {
let is_traversable = |start: &Vec3<f32>, end: &Vec3<f32>| {
vol.ray(*start, *end)
.until(Block::is_solid)
.cast()
.0
.powi(2)
> (*start).distance_squared(*end)
//vol.get(*pos).ok().copied().unwrap_or_else(Block::empty).
// is_fluid();
};
informed_rrt_connect(start, end, is_traversable)
}
}
/// Attempts to find a path from a start to the end using an informed
/// RRT-Connect algorithm. A point is sampled from a bounding spheroid
/// between the start and end. Two separate rapidly exploring random
/// trees extend toward the sampled point. Nodes are stored in k-d trees
/// for quicker nearest node calculations. Points are sampled until the
/// trees connect. A final path is then reconstructed from the nodes.
/// This pathfinding algorithm is more appropriate for 3D pathfinding
/// with wider gaps, such as flying through a forest than for terrain
/// with narrow gaps, such as navigating a maze.
/// Returns a path and whether that path is complete or not.
#[cfg(rrt_pathfinding)]
fn informed_rrt_connect(
start: Vec3<f32>,
end: Vec3<f32>,
is_valid_edge: impl Fn(&Vec3<f32>, &Vec3<f32>) -> bool,
) -> (Option<Path<Vec3<i32>>>, bool) {
let mut path = Vec::new();
// Each tree has a vector of nodes
let mut node_index1: usize = 0;
let mut node_index2: usize = 0;
let mut nodes1 = Vec::new();
let mut nodes2 = Vec::new();
// The parents hashmap stores nodes and their parent nodes as pairs to
// retrace the complete path once the two RRTs connect
let mut parents1 = HashMap::new();
let mut parents2 = HashMap::new();
// The path vector stores the path from the appropriate terminal to the
// connecting node or vice versa
let mut path1 = Vec::new();
let mut path2 = Vec::new();
// K-d trees are used to find the closest nodes rapidly
let mut kdtree1 = KdTree::new();
let mut kdtree2 = KdTree::new();
// Add the start as the first node of the first k-d tree
kdtree1
.add(&[startf.x, startf.y, startf.z], node_index1)
.unwrap_or_default();
nodes1.push(startf);
node_index1 += 1;
// Add the end as the first node of the second k-d tree
kdtree2
.add(&[endf.x, endf.y, endf.z], node_index2)
.unwrap_or_default();
nodes2.push(endf);
node_index2 += 1;
let mut connection1_idx = 0;
let mut connection2_idx = 0;
let mut connect = false;
// Scalar non-dimensional value that is proportional to the size of the
// sample spheroid volume. This increases in value until a path is found.
let mut search_parameter = 0.01;
// Maximum of 7000 iterations
for _i in 0..7000 {
if connect {
break;
}
// Sample a point on the bounding spheroid
let (sampled_point1, sampled_point2) = {
let point = point_on_prolate_spheroid(startf, endf, search_parameter);
(point, point)
};
// Find the nearest nodes to the the sampled point
let nearest_index1 = kdtree1
.nearest_one(
&[sampled_point1.x, sampled_point1.y, sampled_point1.z],
&squared_euclidean,
)
.map_or(0, |n| *n.1);
let nearest_index2 = kdtree2
.nearest_one(
&[sampled_point2.x, sampled_point2.y, sampled_point2.z],
&squared_euclidean,
)
.map_or(0, |n| *n.1);
let nearest1 = nodes1[nearest_index1];
let nearest2 = nodes2[nearest_index2];
// Extend toward the sampled point from the nearest node of each tree
let new_point1 = nearest1 + (sampled_point1 - nearest1).normalized().map(|a| a * radius);
let new_point2 = nearest2 + (sampled_point2 - nearest2).normalized().map(|a| a * radius);
// Ensure the new nodes are valid/traversable
if is_valid_edge(&nearest1, &new_point1) {
kdtree1
.add(&[new_point1.x, new_point1.y, new_point1.z], node_index1)
.unwrap_or_default();
nodes1.push(new_point1);
parents1.insert(node_index1, nearest_index1);
node_index1 += 1;
// Check if the trees connect
if let Ok((check, index)) = kdtree2.nearest_one(
&[new_point1.x, new_point1.y, new_point1.z],
&squared_euclidean,
) {
if check < radius {
let connection = nodes2[*index];
connection2_idx = *index;
nodes1.push(connection);
connection1_idx = nodes1.len() - 1;
parents1.insert(node_index1, node_index1 - 1);
connect = true;
}
}
}
// Repeat the validity check for the second tree
if is_valid_edge(&nearest2, &new_point2) {
kdtree2
.add(&[new_point2.x, new_point2.y, new_point1.z], node_index2)
.unwrap_or_default();
nodes2.push(new_point2);
parents2.insert(node_index2, nearest_index2);
node_index2 += 1;
// Again check for a connection
if let Ok((check, index)) = kdtree1.nearest_one(
&[new_point2.x, new_point2.y, new_point1.z],
&squared_euclidean,
) {
if check < radius {
let connection = nodes1[*index];
connection1_idx = *index;
nodes2.push(connection);
connection2_idx = nodes2.len() - 1;
parents2.insert(node_index2, node_index2 - 1);
connect = true;
}
}
}
// Increase the search parameter to widen the sample volume
search_parameter += 0.02;
}
if connect {
// Construct paths from the connection node to the start and end
let mut current_node_index1 = connection1_idx;
while current_node_index1 > 0 {
current_node_index1 = *parents1.get(&current_node_index1).unwrap_or(&0);
path1.push(nodes1[current_node_index1].map(|e| e.floor() as i32));
}
let mut current_node_index2 = connection2_idx;
while current_node_index2 > 0 {
current_node_index2 = *parents2.get(&current_node_index2).unwrap_or(&0);
path2.push(nodes2[current_node_index2].map(|e| e.floor() as i32));
}
// Join the two paths together in the proper order and remove duplicates
path1.pop();
path1.reverse();
path.append(&mut path1);
path.append(&mut path2);
path.dedup();
} else {
// If the trees did not connect, construct a path from the start to
// the closest node to the end
let mut current_node_index1 = kdtree1
.nearest_one(&[endf.x, endf.y, endf.z], &squared_euclidean)
.map_or(0, |c| *c.1);
// Attempt to pick a node other than the start node
for _i in 0..3 {
if current_node_index1 == 0
|| nodes1[current_node_index1].distance_squared(startf) < 4.0
{
if let Some(index) = parents1.values().choose(&mut thread_rng()) {
current_node_index1 = *index;
} else {
break;
}
} else {
break;
}
}
path1.push(nodes1[current_node_index1].map(|e| e.floor() as i32));
// Construct the path
while current_node_index1 != 0 && nodes1[current_node_index1].distance_squared(startf) > 4.0
{
current_node_index1 = *parents1.get(&current_node_index1).unwrap_or(&0);
path1.push(nodes1[current_node_index1].map(|e| e.floor() as i32));
}
path1.reverse();
path.append(&mut path1);
}
let mut new_path = Vec::new();
let mut node = path[0];
new_path.push(node);
let mut node_idx = 0;
let num_nodes = path.len();
let end = path[num_nodes - 1];
while node != end {
let next_idx = if node_idx + 4 > num_nodes - 1 {
num_nodes - 1
} else {
node_idx + 4
};
let next_node = path[next_idx];
let start_pos = node.map(|e| e as f32 + 0.5);
let end_pos = next_node.map(|e| e as f32 + 0.5);
if vol
.ray(start_pos, end_pos)
.until(Block::is_solid)
.cast()
.0
.powi(2)
> (start_pos).distance_squared(end_pos)
{
node_idx = next_idx;
new_path.push(next_node);
} else {
node_idx += 1;
}
node = path[node_idx];
}
path = new_path;
}
/// Returns a random point within a radially symmetrical ellipsoid with given
/// foci and a `search parameter` to determine the size of the ellipse beyond
/// the foci. Technically the point is within a prolate spheroid translated and
/// rotated to the proper place in cartesian space.
/// The search_parameter is a float that relates to the length of the string for
/// a two dimensional ellipse or the size of the ellipse beyond the foci. In
/// this case that analogy still holds as the ellipse is radially symmetrical
/// along the axis between the foci. The value of the search parameter must be
/// greater than zero. In order to increase the sample area, the
/// search_parameter should be increased linearly as the search continues.
#[allow(clippy::many_single_char_names)]
#[cfg(rrt_pathfinding)]
pub fn point_on_prolate_spheroid(
focus1: Vec3<f32>,
focus2: Vec3<f32>,
search_parameter: f32,
) -> Vec3<f32> {
let mut rng = thread_rng();
// Uniform distribution
let range = Uniform::from(0.0..1.0);
// Midpoint is used as the local origin
let midpoint = 0.5 * (focus1 + focus2);
// Radius between the start and end of the path
let radius: f32 = focus1.distance(focus2);
// The linear eccentricity of an ellipse is the distance from the origin to a
// focus A prolate spheroid is a half-ellipse rotated for a full revolution
// which is why ellipse variables are used frequently in this function
let linear_eccentricity: f32 = 0.5 * radius;
// For an ellipsoid, three variables determine the shape: a, b, and c.
// These are the distance from the center/origin to the surface on the
// x, y, and z axes, respectively.
// For a prolate spheroid a and b are equal.
// c is determined by adding the search parameter to the linear eccentricity.
// As the search parameter increases the size of the spheroid increases
let c: f32 = linear_eccentricity + search_parameter;
// The width is calculated to prioritize increasing width over length of
// the ellipsoid
let a: f32 = (c.powi(2) - linear_eccentricity.powi(2)).powf(0.5);
// The width should be the same in both the x and y directions
let b: f32 = a;
// The parametric spherical equation for an ellipsoid measuring from the
// center point is as follows:
// x = a * cos(theta) * cos(lambda)
// y = b * cos(theta) * sin(lambda)
// z = c * sin(theta)
//
// where -0.5 * PI <= theta <= 0.5 * PI
// and 0.0 <= lambda < 2.0 * PI
//
// Select these two angles using the uniform distribution defined at the
// beginning of the function from 0.0 to 1.0
let rtheta: f32 = PI * range.sample(&mut rng) - 0.5 * PI;
let lambda: f32 = 2.0 * PI * range.sample(&mut rng);
// Select a point on the surface of the ellipsoid
let point = Vec3::new(
a * rtheta.cos() * lambda.cos(),
b * rtheta.cos() * lambda.sin(),
c * rtheta.sin(),
);
// NOTE: Theoretically we should sample a point within the spheroid
// requiring selecting a point along the radius. In my tests selecting
// a point *on the surface* of the spheroid results in sampling that is
// "good enough". The following code is commented out to reduce expense.
//let surface_point = Vec3::new(a * rtheta.cos() * lambda.cos(), b *
// rtheta.cos() * lambda.sin(), c * rtheta.sin()); let magnitude =
// surface_point.magnitude(); let direction = surface_point.normalized();
//// Randomly select a point along the vector to the previously selected surface
//// point using the uniform distribution
//let point = magnitude * range.sample(&mut rng) * direction;
// Now that a point has been selected in local space, it must be rotated and
// translated into global coordinates
// NOTE: Don't rotate about the z axis as the point is already randomly
// selected about the z axis
//let dx = focus2.x - focus1.x;
//let dy = focus2.y - focus1.y;
let dz = focus2.z - focus1.z;
// Phi and theta are the angles from the x axis in the x-y plane and from
// the z axis, respectively. (As found in spherical coordinates)
// These angles are used to rotate the random point in the spheroid about
// the local origin
//
// Rotate about z axis by phi
//let phi: f32 = if dx.abs() > 0.0 {
// (dy / dx).atan()
//} else {
// 0.5 * PI
//};
// This is unnecessary as rtheta is randomly selected between 0.0 and 2.0 * PI
// let rot_z_mat = Mat3::new(phi.cos(), -1.0 * phi.sin(), 0.0, phi.sin(),
// phi.cos(), 0.0, 0.0, 0.0, 1.0);
// Rotate about perpendicular vector in the xy plane by theta
let theta: f32 = if radius > 0.0 {
(dz / radius).acos()
} else {
0.0
};
// Vector from focus1 to focus2
let r_vec = focus2 - focus1;
// Perpendicular vector in xy plane
let perp_vec = Vec3::new(-1.0 * r_vec.y, r_vec.x, 0.0).normalized();
let l = perp_vec.x;
let m = perp_vec.y;
let n = perp_vec.z;
// Rotation matrix for rotation about a vector
let rot_2_mat = Mat3::new(
l * l * (1.0 - theta.cos()),
m * l * (1.0 - theta.cos()) - n * theta.sin(),
n * l * (1.0 - theta.cos()) + m * theta.sin(),
l * m * (1.0 - theta.cos()) + n * theta.sin(),
m * m * (1.0 - theta.cos()) + theta.cos(),
n * m * (1.0 - theta.cos()) - l * theta.sin(),
l * n * (1.0 - theta.cos()) - m * theta.sin(),
m * n * (1.0 - theta.cos()) + l * theta.sin(),
n * n * (1.0 - theta.cos()) + theta.cos(),
);
// Get the global coordinates of the point by rotating and adding the origin
// rot_z_mat is unneeded due to the random rotation defined by lambda
// let global_coords = midpoint + rot_2_mat * (rot_z_mat * point);
midpoint + rot_2_mat * point
}