The one-sentence version
In 2020, during a simulated aerial combat competition organized by DARPA, an AI trained via reinforcement learning defeated an experienced human pilot with a score of 5–0. Between 2023 and 2024, AI has completed aerial combat testing against human-piloted aircraft on a modified real-world experimental aircraft. The discussion sparked by such research goes far beyond "how well the AI flies": how to prove that a learning-based system is safe in high-risk scenarios like flight, and to what extent humans must retain control, are two core issues that aviation AI cannot avoid.
Regarding Timeliness and Scope
This article only compiles publicly reported milestones and the technical and governance issues within them; it has not been cross-checked online item by item.
Layer 1: The Intuition
1.1 Why flight is suitable for studying AI autonomy
- Physical laws are clear, allowing for massive training in high-fidelity simulations;
- Decision-making requires extreme speed, and human reaction speeds have an upper limit;
- Risks are extremely high, so the requirements for safety verification are the strictest.
1.2 Two major difficulties
| Difficulty | Description |
|---|---|
| Simulation to Reality | Strategies trained in simulation encounter sensor noise, air currents, and hardware latency when applied to real aircraft |
| Safety Proof | Traditional flight control software can be reviewed line-by-line to prove behavior; the behavior of neural networks is difficult to verify exhaustively |
Layer 2: Public Milestones
| Time | Event |
|---|---|
| 2020 | DARPA "AlphaDogfight Trials": Multiple teams' AIs competed in simulators; the reinforcement learning agent from the champion (Heron Systems) defeated a human pilot in simulation |
| 2022–2023 | DARPA "Air Combat Evolution" (ACE) program: AI conducted real-aircraft flight tests on a modified X-62A (VISTA) experimental aircraft; a human safety pilot was always on board to take control at any time |
| 2024 | The US Secretary of the Air Force flew on an AI-controlled X-62A for a test flight, drawing public attention |
| Civilian Direction | Autopilots and auto-landing have long been widespread; recent research focuses on the autonomy of single-pilot operations, cargo drones, and Urban Air Mobility (UAM) |
Almost all real-aircraft tests share a commonality: a human safety pilot is on board, and there is an independent safety monitoring system that switches back to manual or conservative mode if the AI's commands exceed safe boundaries.
Layer 3: Technical and Governance Issues
3.1 Runtime assurance
A common safety architecture:
The monitor checks whether the AI's commands will cause the aircraft to enter a dangerous state; if so, it switches to a certified backup controller. This way, there is no need to prove that the AI itself is always correct; one only needs to prove that the monitor and the backup are correct. This is consistent with the logic of why an AI Agent needs a sandbox to "assume it will make mistakes and limit the consequences."
3.2 Certification challenges
Aviation software certification standards (such as DO-178C) require every requirement to be traceable and every code path to be covered. Neural networks have no concept of "code paths"; their behavior is determined by millions of parameters. Aviation regulatory agencies (such as the European Union Aviation Safety Agency, EASA) have already released a roadmap for machine learning applications, but a complete certification system is still under construction.
3.3 The role of humans
Whether in civilian or military contexts, international discussions focus on one question: how much control should humans retain in critical decisions. The consensus in the civil aviation field is to introduce autonomy gradually and in stages; for autonomous systems involving the use of force, it remains an ethical and legal issue under continuous discussion by multilateral frameworks like the United Nations.
Summary
| Layer | Key points |
|---|---|
| Milestones | 2020 simulation combat; 2023–2024 real-aircraft testing, with a safety pilot always on board |
| Difficulties | Simulation to reality; safety proof for learning-based systems |
| Architecture | Runtime assurance: Verifiable monitor + certified backup controller |
| Governance | Certification standards are still under construction; the degree of human control is a core issue |