I Built an AI Pilot That Plans Like a Robot and Dodges Like a Human

2026-05-16 · Source: Towards AI - Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Intermediate, long

Summary

A developer created an AI pilot for an SR-71 aircraft in Unity 2022.3, demonstrating a hybrid architecture that combines classical AI with deep reinforcement learning. The system uses an A* algorithm for routine path planning through a procedural maze and a Deep Q-Network (DQN) for real-time missile evasion. This hybrid agent was trained over 5 million steps using PPO via the ML-Agents trainer, with 24 parallel environments to enhance sample efficiency. The project highlights that classical methods excel at deterministic problems like route planning, while RL is superior for non-stationary, partially observable problems like threat evasion. The architecture allows for a real-time switch between the A* planner and the DQN policy, triggered by a sphere-cast around the missile launcher.

Key takeaway

For AI Engineers developing autonomous systems, consider adopting hybrid architectures that integrate classical algorithms with deep reinforcement learning. This approach can significantly improve sample efficiency and overall system robustness by assigning each problem type (e.g., routine planning vs. dynamic evasion) to the most suitable AI paradigm. Your systems will benefit from the predictability of classical methods and the adaptive capabilities of learned policies.

Key insights

Hybrid AI architectures combining classical and deep learning methods are more effective for complex autonomous systems.

Principles

• Use classical AI for deterministic problems.
• Employ RL for non-stationary, partially observable challenges.
• Stabilize inner-loop dynamics with classical control.

Method

The system switches between an A* path planner for routine navigation and a DQN policy for threat evasion, triggered by real-time environmental cues like missile detection. Training uses 24 parallel Unity environments for sample efficiency.

In practice

• Implement A* for known obstacle avoidance.
• Train DQN for dynamic threat evasion.
• Use parallel environments for faster RL training.

Topics

• Hybrid AI Architecture
• Deep Reinforcement Learning
• A* Pathfinding
• Unity ML-Agents
• Autonomous Systems

Code references

• Alpsource/Mastering-Fighter-Jet-Survival-Tactics

Best for: AI Engineer, Machine Learning Engineer, Robotics Engineer

Related on AIssential

• Why the Next Leap in AI May Depend on Learning by Trial and Error — Deep reinforcement learning offers AI a crucial mechanism for learning through action and discovery, moving beyond mere pattern recognition.
• From Kinematics to Dynamics: Learning to Refine Hybrid Plans for Physically Feasible Execution — Robotic plans based on first-order dynamics are physically infeasible; a neuro-symbolic RL framework can refine them for second-order feasibility.
• Dertouzos Distinguished Lecturer: Richard Sutton — OAK proposes a domain-independent AI architecture that learns abstractions and skills from experience, scaling with computation.
• Lessons from Trillion Token Deployments at Fortune 500s — Alessandro Cappelli, Adaptive ML — Reinforcement Learning systematically integrates feedback to accelerate LLM lifecycle, enabling production-ready, cost-effective, and performant models.
• Reinforcement Learning - Ep. 30 (Deep Learning SIMPLIFIED) — Reinforcement learning enables agents to maximize rewards by learning optimal actions within dynamic, uncertain environments.
• From Next-Word Prediction to World Models — AI is evolving from single-model next-word prediction to complex, efficient, and agentic "world models."

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards AI - Medium.