A simulation environment inspired by OpenAI Gym and PettingZoo that explores autonomous coordination between two lunar landers attempting simultaneous touchdowns. This project focuses on extending the traditional single-agent Lunar Lander task to a multi-agent reinforcement learning (MARL) scenario — introducing complexity, real-time cooperation, collision avoidance, and fuel optimization.
Multi-Agent Lunar Lander is an advanced version of the classical Lunar Lander environment built using OpenAI Gymnasium (Box2D) and the PettingZoo framework. Instead of controlling a single lander, this project introduces two landers that must coordinate to land safely — transforming the task from a single-agent control problem into a multi-agent coordination challenge.
Agents must manage:
- Independent and interdependent controls
- Thrust, orientation, fuel optimization
- Inter-agent interference and stability management
The simulation resembles real-world scenarios such as:
- Coordinated landings of autonomous spacecraft
- Drone fleet coordination
- Multi-robot system control in unstructured environments
In future space missions, simultaneous landings of multiple landers on the Moon could help reduce costs via:
- Shared payload capacity
- Rideshare opportunities
- Reusability
However, this approach increases:
- Navigational complexity
- Development cost for hazard avoidance and control systems
- Risk of interference and failure
This simulation environment allows for experimentation and training of agents under such constraints, enabling researchers to optimize for cost, safety, and performance.
The project’s objectives include:
- Simulating safe and fuel-efficient landings in a multi-agent scenario
- Designing adaptive control strategies using reinforcement learning
- Building agents that generalize across dynamic conditions and unforeseen environmental states
- Benchmarking various RL algorithms using performance metrics such as:
- Landing success
- Fuel efficiency
- Collision avoidance
- Time-to-land
- Cost Efficiency: Reduces need for redundant backup systems
- Mission Reliability: Promotes robust risk assessment and coordination
- Improved Precision: Helps test real-time autonomous landing under limited zone constraints
- Scalability: Encourages multi-lander, multi-mission automation
Synchronization between landers is crucial — failure in coordination can increase mission time and operational costs.
This project:
- Converts OpenAI’s single-agent Lunar Lander into a multi-agent PettingZoo-compatible environment
- Simulates real-world lunar dynamics such as:
- Irregular terrain
- Varied soil types
- Light and gravity conditions
- Dual-agent control with independent and shared policy learning
- Custom reward functions for balancing:
- Safe landing
- Fuel use
- Synchronization
- Collision avoidance
- Parameter tuning for different operation scenarios
- Modular design for research extensibility and reproducibility
OpenAI GymPettingZooBox2DStable-Baselines3Python 3.xNumPy,Matplotlib, etc.
Inspired by NASA's CADRE and AAMAS projects, which explore decentralized coordination among autonomous robotic landers.
We welcome contributions! Potential extensions:
- Inter-agent communication modeling
- Competitive vs cooperative multi-agent settings
- Integration with real-world sensor data
This project is open-sourced under the MIT License.
Multi-Agent Lunar Lander provides a challenging yet promising platform for advancing reinforcement learning in space robotics, autonomous control, and multi-agent systems. Through research and collaboration, this project aims to serve as a testbed for future intelligent space missions.