dodgy

A Rust crate to compute local collision avoidance (specifically ORCA) for agents.

Why local collision avoidance?

Characters in video games generally need to find paths to navigate around the game world. Once this is done, the path needs to be followed. The trouble occurs when characters start getting in the way of each other. As paths are not generally regenerated every game frame, other characters cannot be taken into account. Local collision avoidance provides cheap avoidance for characters even in high-density situations.

Which local collision avoidance?

There are several algorithms for local collision avoidance. This crate implements ORCA.

This crate is essentially a port of RVO2 to Rust. Several changes have been made: tests have been written, code more commented, and the public API made more flexible.

Example

This example uses the "raw" API.

use dodgy::{Agent, AvoidanceOptions, Obstacle};
use glam::Vec2;

let mut agents = vec![
  Agent {
    position: Vec2::ZERO,
    velocity: Vec2::ZERO,
    radius: 1.0,
    avoidance_responsibility: 1.0,
    max_velocity: 5.0,
  },
  // Add more agents here.
];

let goal_points = vec![
  Vec2::new(50.0, 0.0),
  // Add goal points for every agent.
];

let obstacles = vec![
  Obstacle::Closed{
    vertices: vec![
      Vec2::new(-1000.0, -1000.0),
      Vec2::new(-1000.0, 1000.0),
      Vec2::new(1000.0, 1000.0),
      Vec2::new(1000.0, -1000.0),
    ],
  },
  // Add more obstacles here.
];

let time_horizon = 3.0;
let obstacle_time_horizon = 1.0;

fn get_delta_seconds() -> f32 {
  // Use something that actually gets the time between frames.
  return 0.01;
}

for i in 0..100 {
  let delta_seconds = get_delta_seconds();
  if delta_seconds == 0.0 {
    // Skip frames where agents can't move anyway.
    continue;
  }

  let mut new_velocities = Vec::with_capacity(agents.len());

  for i in 0..agents.len() {
    let neighbours = agents[..i]
      .iter()
      .chain(agents[(i + 1)..].iter())
      .collect::<Vec<&Agent>>();
    let nearby_obstacles = obstacles
      .iter()
      .map(|obstacle| obstacle)
      .collect::<Vec<&Obstacle>>();

    let preferred_velocity = (goal_points[i] - agents[i].position)
      .normalize_or_zero() * agents[i].max_velocity;

    let avoidance_velocity = agents[i].compute_avoiding_velocity(
      &neighbours,
      &nearby_obstacles,
      preferred_velocity,
      delta_seconds,
      &AvoidanceOptions {
        obstacle_margin: 0.1,
        time_horizon,
        obstacle_time_horizon,
      });
    new_velocities.push(avoidance_velocity);
  }

  for (i, agent) in agents.iter_mut().enumerate() {
    agent.velocity = new_velocities[i];
    agent.position += agent.velocity * delta_seconds;
  }

  // Update rendering using new agent positions.
}

This is the preferred API to use, since finding neighbours is essentially just a spatial query. Often, finding related objects within some radius is performed anyway in most game engines. Using this API, neighbours can be find through your regular spatial queries, and exposes just the avoidance part.

However, an alternative using the Simulator struct:

use dodgy::{Agent, AvoidanceOptions, AgentParameters, Obstacle, Simulator};
use glam::Vec2;

let mut simulator = Simulator::new();
simulator.add_agent(Agent {
  position: Vec2::ZERO,
  velocity: Vec2::ZERO,
  radius: 1.0,
  avoidance_responsibility: 1.0,
  max_velocity: 5.0,
}, AgentParameters {
  goal_point: Vec2::new(50.0, 0.0),
  obstacle_margin: dodgy::SimulatorMargin::Distance(0.1),
  time_horizon: 3.0,
  obstacle_time_horizon: 1.0,
});
// Add more agents.

simulator.add_obstacle(
  Obstacle::Closed{
    vertices: vec![
      Vec2::new(-1000.0, -1000.0),
      Vec2::new(-1000.0, 1000.0),
      Vec2::new(1000.0, 1000.0),
      Vec2::new(1000.0, -1000.0),
    ],
  }
);
// Add more obstacles.

fn get_delta_seconds() -> f32 {
  // Use something that actually gets the time between frames.
  return 0.01;
}

for i in 0..100 {
  let delta_seconds = get_delta_seconds();
  simulator.step(delta_seconds);

  // Update rendering using new agent positions (using simulator.get_agent).
}

Again, this is not the preferred method! It is just a simpler way to get up and running for a small group of users. The other API is more flexible and preferred.

License

Licensed under the MIT license.

Attribution

dodgy contains code ported from RVO2. See original_license.txt.