Struct EnvPool 

pub struct EnvPool<E: Environment> { /* private fields */ }

A pool of environments for parallel execution

EnvPool manages multiple environment instances and executes operations across them in parallel using Rayon’s thread pool. This provides significant performance improvements over sequential execution.

Performance

For N environments with average step time T:

• Sequential: O(N * T)
• Parallel: O(max(T)) ≈ O(T) when N ≤ num_cores

This can provide 10-100x speedups depending on environment complexity and number of CPU cores available.

Action type

EnvPool is currently restricted to discrete-action envs (E::Action = i64). The pool’s step signature passes &[i64] because every existing trainer and example expects a dense vector of integer action indices. Adding pool support for continuous-action envs (e.g. SAC) is a separate follow-up; see issue #61 for the rationale behind the associated-type design.

Implementations

impl<E: Environment<Action = i64> + Send> EnvPool<E>

pub fn new<F>(env_fn: F, num_envs: usize) -> Self

where F: Fn() -> E,

Create a new environment pool

Arguments

• env_fn - Factory function to create environment instances
• num_envs - Number of parallel environments

Example

use thrust_rl::env::{cartpole::CartPole, pool::EnvPool};

let pool = EnvPool::new(|| CartPole::new(), 8);

pub fn reset(&mut self) -> Vec<Vec<f32>>

where E: Send,

Reset all environments in parallel

Returns a vector of initial observations, one per environment.

Example

let observations = pool.reset();
assert_eq!(observations.len(), 4);

pub fn step(&mut self, actions: &[i64]) -> Vec<StepResult>

where E: Send,

Step all environments in parallel with given actions

Arguments

• actions - Slice of actions, one per environment

Panics

Panics if the number of actions doesn’t match the number of environments.

Example

let actions = vec![0, 1, 0, 1];
let results = pool.step(&actions);
assert_eq!(results.len(), 4);

pub fn num_envs(&self) -> usize

Get the number of environments in the pool

pub fn observation_space(&self) -> SpaceInfo

Get observation space information from first environment

pub fn action_space(&self) -> SpaceInfo

Get action space information from first environment

pub fn reset_env(&mut self, env_id: usize) -> Result<Vec<f32>>

Reset a specific environment by index

Arguments

• env_id - Index of environment to reset

Returns

Initial observation from the reset environment

impl<E: Environment<Action = i64> + Send> EnvPool<E>

pub fn step_structured(&mut self, actions: &[i64]) -> PoolStepResult<Vec<f32>>

Step all environments and return structured result

This is a convenience method that unpacks individual StepResults into a single PoolStepResult with parallel vectors.