Operant

High-performance SIMD-optimized Gymnasium-compatible reinforcement learning environments in Rust with Python bindings.

~600x faster than Gymnasium for vectorized environments.

What is This?

Operant provides native Rust implementations of Gymnasium environments with:

• SIMD vectorization: Process 8 environments simultaneously per instruction (AVX2)
• Struct-of-Arrays layout: Cache-friendly memory access patterns
• Zero-copy numpy: Direct array access without Python overhead
• Gymnasium compatibility: Drop-in replacement for standard Gym environments

Unlike PufferLib which wraps existing Gymnasium environments for vectorization, Operant implements environments natively in Rust for maximum performance.

Supported Environments

All environments provide Gymnasium-compatible observation_space and action_space properties for easy integration with RL frameworks.

Performance

CartPole Benchmark (4096 envs)
============================================================
Operant...     97.54M steps/sec
Gymnasium...    0.16M steps/sec

Speedup: ~600x faster than Gymnasium

Requirements

• Python 3.10+

Installation

Python (PyPI)

pip install operant

Rust (crates.io)

cargo add operant

From Source (Development)

Requires Rust nightly and Poetry:

# 1. Install Rust
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh

# 2. Install Poetry
curl -sSL https://install.python-poetry.org | python3 -

# 3. Setup project
poetry install
poetry run maturin develop --release

Usage

Python

CartPole (Discrete Actions)

import numpy as np
from operant.envs import CartPoleVecEnv

# Create 4096 parallel environments
num_envs = 4096
env = CartPoleVecEnv(num_envs)
obs, info = env.reset(seed=42)  # Shape: (4096, 4)

for step in range(10000):
    actions = np.random.randint(0, 2, size=num_envs, dtype=np.int32)
    obs, rewards, terminals, truncations, info = env.step(actions)

Multi-threaded Execution

For heavier environments or large batch sizes, enable parallel execution:

# Use 4 worker threads for parallel step execution
env = CartPoleVecEnv(num_envs=8192, workers=4)

MountainCar (Discrete Actions)

from operant.envs import MountainCarVecEnv

num_envs = 4096
env = MountainCarVecEnv(num_envs)
obs, info = env.reset(seed=42)  # Shape: (4096, 2)

for step in range(10000):
    actions = np.random.randint(0, 3, size=num_envs, dtype=np.int32)
    obs, rewards, terminals, truncations, info = env.step(actions)

Pendulum (Continuous Actions)

from operant.envs import PendulumVecEnv

num_envs = 4096
env = PendulumVecEnv(num_envs)
obs, info = env.reset(seed=42)  # Shape: (4096, 3) - [cos(θ), sin(θ), θ_dot]

for step in range(10000):
    actions = np.random.uniform(-2.0, 2.0, size=num_envs).astype(np.float32)
    obs, rewards, terminals, truncations, info = env.step(actions)

Rust

use operant::{CartPole, VecEnv};

fn main() {
    // Create 1024 parallel environments
    let mut env = CartPole::new(1024);

    // Reset all environments
    let obs = env.reset();

    // Step with actions
    let actions = vec![0; 1024];
    let (obs, rewards, terminals, truncations) = env.step(&actions);
}

Logging and Metrics

from operant.utils import Logger

# Context manager automatically handles cleanup
with Logger(csv_path="training.csv") as logger:
    for step in range(1000):
        # ... training loop ...
        logger.log(steps=num_envs, reward=mean_reward, length=mean_length)

Migration from v0.1.x

Old imports (deprecated):

from operant import PyCartPoleVecEnv, Logger

New imports (recommended):

from operant.envs import CartPoleVecEnv
from operant.utils import Logger

The old import style will continue to work until v0.4.0, but will emit deprecation warnings.

Benchmarks

Quick Benchmark

Compare Operant at 4096 environments:

poetry run python benches/cartpole_benchmark.py

Full Benchmark

Test across multiple environment counts (1, 16, 256, 1024, 4096):

poetry run python benches/cartpole_benchmark.py --all

Architecture

Operant uses a Struct-of-Arrays (SoA) memory layout with SIMD vectorization:

• f32x8 SIMD: Processes 8 environments simultaneously per instruction
• SoA Layout: Cache-friendly memory access patterns
• Zero-copy: Direct numpy array access without Python overhead
• Rust + PyO3: Native performance with Python ergonomics

Development

# Run tests
poetry run pytest

# Build in debug mode (faster compilation)
poetry run maturin develop

# Build in release mode (faster runtime)
poetry run maturin develop --release

License

MIT