Project Armstrong: AI Agents for Lunar Landing By Deependra Shekhawat

Overview

This project implements a multi-agent AI system to pilot a Lunar Lander safely to the surface. It uses Google's Gemini 2.5 Flash-Lite model via the Agent Development Kit (ADK) to power two distinct agents: a Navigator and a Commander.

This project is designed for the Kaggle Agents Intensive Capstone.

Mission Replay

Watch the AI agents navigate and land the lunar lander in real-time.

https://github.com/dpshekhawat/project-armstrong/blob/main/mission_replay.mp4

mission_replay_sound_edit.mp4

Architecture

For detailed architecture documentation, see ARCHITECTURE.md.

High-Level Design

The system uses a hierarchical multi-agent architecture with ADK components:

1. Navigator Agent (LlmAgent):

   • Role: Strategic Advisor.
   • Input: Raw telemetry (Altitude, Velocity, Angle).
   • Memory: Maintains session-based conversation history via InMemorySessionService.
   • Context Management: Uses EventsCompactionConfig to prevent unbounded history growth.
   • Output: High-level advice (e.g., "Stabilize angle before descending").

2. Commander Agent (LlmAgent):

   • Role: Pilot / Executor.
   • Input: Telemetry + Navigator's Advice.
   • Mechanism: Uses Function Calling (execute_maneuver tool) to interact with the environment.
   • Output: Specific actions (Main Engine, Side Engines) and durations.

3. ADK Integration:

   • Runner: Orchestrates agent execution with session management
   • App: Wraps Navigator with context compaction configuration
   • SessionService: Manages conversation state across turns
   • Gemini Model: Powered by gemini-2.5-flash-lite with retry logic

Key Features

• Multi-Agent System: Hierarchical coordination between Navigator (strategy) and Commander (tactics)
• ADK Framework: Built using Google's Agent Development Kit following course best practices
• Session Management: Navigator maintains conversation history with context compaction
• Function Calling: Commander uses ADK function tools for environment interaction
• Observability: Tracks token usage, latency, and errors for every LLM call
• Robustness: Implements exponential backoff retry logic for API stability
• Physics-Aware Prompts: Agents are primed with specific physics rules (e.g., "Stabilize First") to prevent common failure modes like the "Death Spiral"
• Evaluation: Includes a script to run multiple episodes and calculate metrics

Files

• agents.py: Defines agent creation functions (create_navigator_agent, create_commander_agent) using ADK's LlmAgent
• main_mission.py: Runs a single demo mission with video recording using ADK Runner
• evaluate_agent.py: Runs a batch of episodes to calculate metrics with ADK patterns
• lunar_tools.py: Interface for the Gymnasium LunarLander-v2 environment
• ARCHITECTURE.md: Detailed architecture documentation with diagrams

How to Run

Prerequisites

• Python 3.10+ (recommended) or Python 3.9 (with warnings)
• Conda environment manager
• Visual Studio C++ Build Tools (required for Box2D compilation on Windows)
  • Download from: https://visualstudio.microsoft.com/visual-cpp-build-tools/
  • Or install via: winget install Microsoft.VisualStudio.2022.BuildTools

1. Setup Environment with Conda

# Create conda environment with Python 3.10+
conda create -n armstrong_env python=3.10
conda activate armstrong_env

# Install dependencies
pip install -r requirements.txt

2. Install Box2D (if needed)

# Try this first - requires Visual Studio Build Tools
pip install gymnasium[box2d]

# If compilation fails, install Build Tools then retry

3. Set API Key

Set your GOOGLE_API_KEY in a .env file or environment variable:

# In PowerShell
$env:GOOGLE_API_KEY='your_key_here'

# Or create .env file with:
GOOGLE_API_KEY=your_key_here

4. Run a Mission

# Make sure you're in the conda environment
python main_mission.py

5. Run Evaluation

python evaluate_agent.py

Implementation Notes

ADK Integration Status

✅ Completed:

• Agents use LlmAgent from google.adk.agents
• Session management via InMemorySessionService
• Context compaction with EventsCompactionConfig (10 turns, 2 overlap)
• Function tools following ADK best practices
• Runner orchestration for Navigator agent
• Async/await patterns for ADK compatibility

⚠️ Known Issues:

• Box2D requires Visual Studio C++ Build Tools on Windows
• Python 3.9 shows deprecation warnings (upgrade to 3.10+ recommended)

Rate Limiting

The code includes 9-second delays between steps to respect Gemini API rate limits (15 RPM for free tier with 2 calls per step).

Results

The agents demonstrate the ability to:

• Recover from unstable initial conditions.
• Coordinate strategy (Navigator) and tactics (Commander).
• Land safely (Reward > 200) in successful episodes.