Plangym

Plangym is a Python library that extends Gymnasium environments for planning algorithms. It provides the ability to get and set complete environment state, enabling deterministic rollouts from arbitrary states—critical for planning algorithms that need to branch execution.

Key Features

• State manipulation: get_state() and set_state() for full environment state control
• Batch stepping: Execute multiple state-action pairs in a single call
• Parallel execution: Built-in multiprocessing and Ray support for distributed rollouts
• Gymnasium compatible: Works with gym.Wrappers and standard Gym API
• Delayed initialization: Serialize environments before setup for distributed workers

---

Table of Contents

• Supported Environments
• Requirements
• Installation
• Quick Start
• Developer Guide
• Local CI with act
• Architecture
• License
• Contributing

---

Supported Environments

Environment Type	Package	Description
Classic Control	gymnasium	CartPole, Pendulum, MountainCar, etc.
Box2D	gymnasium[box2d]	LunarLander, BipedalWalker, CarRacing
Atari	ale-py	Atari 2600 games via Arcade Learning Environment
dm_control	dm-control	DeepMind Control Suite with MuJoCo physics
MuJoCo	mujoco	MuJoCo physics environments
Retro	stable-retro	Classic console games (Genesis, SNES, etc.)
NES	nes-py	NES games including Super Mario Bros

---

Requirements

Python Version

• Python 3.10 or higher

System Dependencies

Ubuntu / Debian

# Install all system dependencies for headless rendering (EGL, GLU, X11)
make install-system-deps
# Or manually:
sudo apt-get update
sudo apt-get install -y xvfb libglu1-mesa libegl1-mesa-dev libgl1-mesa-glx x11-utils

For NES environments (nes-py):

sudo apt-get install -y build-essential clang libstdc++-10-dev

macOS

brew install --cask xquartz
brew install swig libzip
# Create X11 socket directory if neededif [ !-d /tmp/.X11-unix ];then
sudo mkdir /tmp/.X11-unix
sudo chmod 1777 /tmp/.X11-unix
sudo chown root /tmp/.X11-unix
fi

WSL2 (Windows)

# Install all system dependencies for headless rendering (EGL, GLU, X11)
make install-system-deps
# Or manually:
sudo apt-get update
sudo apt-get install -y xvfb libglu1-mesa libegl1-mesa-dev libgl1-mesa-glx x11-utils

For GUI rendering, install an X server on Windows (e.g., VcXsrv) or use headless mode.

---

Installation

Quick Install

# Using pip
pip install plangym
# Using uv
uv add plangym

Install with Optional Extras

Plangym provides optional extras for different environment types:

Extra	Description	Includes
atari	Atari 2600 games	ale-py, gymnasium[atari]
nes	NES / Super Mario	nes-py, gym-super-mario-bros
classic-control	Classic control envs	gymnasium[classic_control], pygame
dm_control	DeepMind Control Suite	mujoco, dm-control
retro	Retro console games	stable-retro
box_2d	Box2D physics	box2d-py
ray	Distributed computing	ray
jupyter	Notebook support	jupyterlab

# Install specific extras
pip install "plangym[atari,dm_control]"# Install all environment extras
pip install "plangym[atari,nes,classic-control,dm_control,retro,box_2d,ray]"

Development Installation

git clone https://github.com/FragileTech/plangym.git
cd plangym
uv sync --all-extras

ROM Installation

For Retro environments, you need to import ROM files:

# Retro ROMs (requires ROM files)
python -m plangym.scripts.import_retro_roms

Note: Atari ROMs are now bundled with ale-py >= 0.9, so no additional installation is needed for Atari environments.

---

Quick Start

Basic Environment Stepping

importplangymenv=plangym.make(name="CartPole-v1")
state, obs, info=env.reset()
# Save state for latersaved_state=state.copy()
# Take a stepaction=env.action_space.sample()
new_state, obs, reward, terminated, truncated, info=env.step(state=state, action=action)
# Restore to saved state and try a different actiondifferent_action=env.action_space.sample()
new_state2, obs2, reward2, _, _, _=env.step(state=saved_state, action=different_action)

Batch Stepping

Execute multiple state-action pairs efficiently:

importplangymenv=plangym.make(name="CartPole-v1")
state, obs, info=env.reset()
# Create batch of states and actionsstates= [state.copy() for_inrange(10)]
actions= [env.action_space.sample() for_inrange(10)]
# Step all at oncenew_states, observations, rewards, terminateds, truncateds, infos=env.step_batch(
states=states,
actions=actions
)

Parallel Execution

Use multiple workers for faster rollouts:

importplangym# Create environment with 4 parallel workersenv=plangym.make(name="ALE/MsPacman-v5", n_workers=4)
state, obs, info=env.reset()
states= [state.copy() for_inrange(100)]
actions= [env.action_space.sample() for_inrange(100)]
# Steps are distributed across workersnew_states, observations, rewards, terminateds, truncateds, infos=env.step_batch(
states=states,
actions=actions
)

---

Developer Guide

Development Setup

git clone https://github.com/FragileTech/plangym.git
cd plangym
uv sync --all-extras

Code Style

Plangym uses Ruff for linting and formatting.

# Auto-fix and format code
make style
# Check code style (no modifications)
make check

Running Tests

# Run full test suite
make test# Run tests in parallel (default: 2 workers)
make test-parallel
# Run tests with custom worker count
n=4 make test-parallel
# Run classic control tests (single-threaded)
make test-singlecore
# Run doctests
make test-doctest

Running individual test files:

# dm_control tests (requires MUJOCO_GL for headless rendering)
MUJOCO_GL=egl uv run pytest tests/control/test_dm_control.py -s
# Specific test
uv run pytest tests/test_core.py::TestCoreEnv::test_step -v

Environment Variables:

Variable	Description
MUJOCO_GL=egl	Headless MuJoCo rendering
PYVIRTUALDISPLAY_DISPLAYFD=0	Virtual display for rendering tests
SKIP_CLASSIC_CONTROL=1	Skip classic control in parallel runs
SKIP_RENDER=True	Skip rendering tests
n=2	Number of parallel test workers

Code Coverage

# Run all coverage targets
make codecov
# Individual coverage targets
make codecov-parallel # Parallel tests
make codecov-singlecore # Single-core tests

Building Documentation

# Build Sphinx documentation
make build-docs
# Serve documentation locally
make serve-docs

Docker

# Build Docker image
make docker-build
# Run interactive shell in container
make docker-shell
# Run tests in Docker
make docker-test
# Run Jupyter notebook server
make docker-notebook

---

Local CI with act

act allows you to run GitHub Actions workflows locally for debugging.

Prerequisites

1. Docker installed and running

   • For WSL2: Enable Docker Desktop WSL Integration in Settings → Resources → WSL Integration

2. act installed:

   # macOS
   brew install act
   # Linux (using Go)
   go install github.com/nektos/act@latest
   # Or download from GitHub releases

Configuration

Plangym includes pre-configured act settings:

• .actrc - Default act configuration
• .secrets - Local secrets file (gitignored)

Running Workflows Locally

# List all available jobs
act -l
# Run specific jobs
act -j style-check # Lint check
act -j pytest # Run tests
act -j build-test-package # Build and test package# Dry run (see what would execute)
act -n
# Run with verbose output
act -j style-check -v

Secrets Setup

Edit .secrets to add your credentials for full CI functionality:

# .secrets file format
ROM_PASSWORD=your_rom_password
CODECOV_TOKEN=your_codecov_token
TEST_PYPI_PASS=your_test_pypi_token
BOT_AUTH_TOKEN=your_github_bot_token
PYPI_PASS=your_pypi_token

Note: The .secrets file is gitignored and should never be committed.

Troubleshooting act

Docker not found in WSL2

Enable WSL integration in Docker Desktop:

1. Open Docker Desktop
2. Go to Settings → Resources → WSL Integration
3. Enable integration for your WSL distro
4. Restart Docker Desktop

Job runs but fails on specific actions

Some GitHub Actions may not work perfectly with act. Common issues:

• actions/cache - May need --reuse flag
• Platform-specific steps - act only runs Linux containers
• Service containers - May require additional configuration

---

Architecture

plangym/
├── core.py # PlanEnv, PlangymEnv base classes
├── registry.py # make() factory function
├── control/ # Physics environments
│ ├── classic_control.py
│ ├── dm_control.py
│ ├── mujoco.py
│ └── box2d.py
├── videogames/ # Emulator environments
│ ├── atari.py
│ ├── retro.py
│ └── nes.py
└── vectorization/ # Parallel execution
├── env.py # VectorizedEnv base
├── parallel.py # Multiprocessing
└── ray.py # Ray distributed

Core Classes

Class	Description
PlanEnv	Abstract base defining get_state(), set_state(), step() interface
PlangymEnv	Wraps Gymnasium environments with state manipulation
VectorizedEnv	Base for parallel execution backends
ParallelEnv	Multiprocessing-based parallel stepping
RayEnv	Ray-based distributed stepping

Entry Point

importplangym# The make() function routes to the correct environment classenv=plangym.make(
name="CartPole-v1", # Environment namen_workers=4, # Parallel workers (optional)obs_type="rgb", # Observation type: coords, rgb, grayscaledelay_setup=True, # Defer initialization for serialization
)

---

License

Plangym is released under the MIT License.

---

Contributing

Contributions are welcome! Please read our Contributing Guidelines before submitting a pull request.

Quick contribution checklist:

1. Run make check to verify code style
2. Run make test to ensure tests pass
3. Add tests for new functionality
4. Update documentation as needed

For bug reports and feature requests, please open an issue.