Multi-Agent RL MAPF Drone Demo

Instant-alignment drone AI for safe, adaptive, real-time flight control

Overview

Instead of waiting for delayed success signals, this system scores every action instantly against a set of internal principles (the “Constitution”) so it can act safely and efficiently in real time.

Drone Reinforcement Learning (RL) and Multi-Agent Path Finding (MAPF) Alignment System – Complete Overview
This drone artificial intelligence system runs on three core agents that sense, process, and decide actions in real time using internal values rather than delayed rewards. A Safety Controller blocks risky moves, and a Supervisor keeps every component running smoothly and safely.

How the System Works

The program simulates a distributed control system for a drone, built around three independent agents:

• Ingestion Agent
  Reads raw sensor data such as camera feeds, global positioning system (GPS) readings, inertial measurement unit (IMU) readings, and barometer values at a fixed frequency. Pushes the data into bounded queues for processing.

• Preprocess Agent
  Cleans, filters, and compresses incoming data into a compact set of features that are easier for the decision-making process to use.

• Prediction Agent
  Uses the processed features to select the drone’s next action (hover, climb, turn left/right, move forward).

Alignment Scoring

Instead of delayed rewards, each action is scored immediately based on internal values such as:

• Maintain safety margins
• Save energy
• Move smoothly

Each value returns a score between 0 and 1, weighted by importance. Scores are combined and tested with small variations (structural inversion test) to ensure stability. The Alignment Policy chooses the action with the highest current alignment score to prevent unsafe or short-sighted decisions.

Safety and Oversight

• Safety Controller
  Runs outside the main AI path and can override any decision if it detects danger (for example, entering a no-fly zone).
• Supervisor
  Starts all agents and the Safety Controller, monitors performance, restarts failed agents, and quarantines components that fail repeatedly.

Operational Characteristics

• Runs continuously and adapts to changing environmental conditions.
• Uses continuous alignment scoring rather than delayed reinforcement learning.
• Combines real-time decision making with hard safety overrides.

System Design Files

• drone_high_lv_system_design.png – High-level diagram showing drone agents, Safety Controller, Supervisor, server components, and data/control flow.
• drone_low_lv_system_design.png – Low-level diagram showing process flows, interfaces, and message passing.

Core Code

• drone_main.py – Main program entry point. Initializes the Supervisor, starts agents, manages the Safety Controller, and handles alignment scoring.

Reference Material

• drones_matrix_RL.png – Diagram comparing vertical vs. horizontal reward patterns in drone RL, showing trade-offs between event-based rewards and continuous alignment scoring.

Design Document

• low_level_design – File containing detailed subsystem designs, flowcharts, and specifications for each system module.