This example demonstrates a visual Gymnasium environment (LunarLander) integrated with the Model Context Protocol (MCP), featuring image rendering and complex dependencies that test conda isolation.

The LunarLander environment is a classic rocket trajectory optimization problem where an agent must learn to land a spacecraft safely on a landing pad. This example serves as a test case for:

• Visual Environment Support: Returns rendered frames as base64-encoded images
• Complex Dependencies: Requires swig and gymnasium[box2d] for box2d physics
• Conda Isolation Testing: Tests the managed simulation server's ability to handle environments with external dependencies

This example requires external dependencies that must be properly handled by conda isolation:

# Required system dependency
pip install swig

# Gymnasium with box2d physics
pip install gymnasium[box2d]

# Rendering support
pip install pygame

• Action Space: Discrete(4) - NOTHING, FIRE_LEFT, FIRE_MAIN, FIRE_RIGHT
• Observation Space: Box(8) - [x, y, vx, vy, angle, angular_velocity, leg1_contact, leg2_contact]
• Reward: Based on distance to landing pad, velocity, angle, fuel consumption, and landing success
• Episode End: Landing (success/crash), going out of bounds, or time limit

cd examples/lunar_lander_mcp/mcp_server
python lunar_lander_mcp_server.py --port 8000 --seed 42

cd examples/frozen_lake_mcp_complete/mcp_server
python managed_simulation_server.py --port 9002 --use-conda-isolation \
  --production-script /path/to/lunar_lander_mcp_server.py \
  --requirements /path/to/requirements.txt

cd examples/lunar_lander_mcp
python generate_sample_images.py

This creates a sample_trajectory/ directory with:

• step_*.png - Rendered frames showing the lander in action
• trajectory_summary.json - Complete trajectory data with observations and rewards

cd examples/lunar_lander_mcp
python test_lunar_lander_conda.py

This verifies that the managed simulation server can properly:

• Create isolated conda environments
• Install complex dependencies (swig, box2d)
• Run lunar lander simulations with visual rendering

• game://initial_state - Initial environment state with rendered frame
• game://current_frame - Current rendered frame as base64 image
• game://action_space - Available actions and descriptions
• game://observation_space - Observation vector description

• lander_action - Control the lunar lander with actions: NOTHING, FIRE_LEFT, FIRE_MAIN, FIRE_RIGHT

This example is specifically designed to test conda environment isolation:

1. Complex Dependencies: Requires compilation of box2d via swig
2. Visual Rendering: Tests pygame integration
3. Dependency Management: Verifies requirements.txt installation in isolated environments

To test conda isolation:

# This should create a fresh conda environment and install all dependencies
python managed_simulation_server.py --port 9002 --use-conda-isolation --verbose

• Successful Landing: Lander touches down gently with both legs, positive reward
• Crash: Lander hits ground too hard or at wrong angle, negative reward
• Out of Bounds: Lander leaves the visible area, episode terminates
• Visual Feedback: Each action returns a rendered frame showing the current state

lunar_lander_mcp/
├── README.md                     # This file
└── mcp_server/
    ├── requirements.txt          # Dependencies for conda isolation
    ├── lunar_lander_adapter.py   # Environment adapter with rendering
    └── lunar_lander_mcp_server.py # MCP server implementation

If you encounter swig compilation errors:

# Ubuntu/Debian
sudo apt-get install swig

# macOS
brew install swig

# Windows
# Download from http://www.swig.org/download.html

# Clear pip cache and reinstall
pip cache purge
pip install --no-cache-dir gymnasium[box2d]

# Install pygame if rendering fails
pip install pygame

# For headless environments, you may need virtual display
sudo apt-get install xvfb
export DISPLAY=:99
Xvfb :99 -screen 0 1024x768x24 &