Is AI Ready to Fight Wars?

Artificial intelligence is making remarkable strides in fields previously dominated by humans, with recent developments highlighting its potential in both military and recreational applications. One notable advancement comes from DARPA’s Air Combat Evaluation (ACE) program, which has successfully tested an AI pilot in a jet fighter against human opponents. This achievement underscores the ongoing effort to develop aerospace AI agents capable of performing safely and autonomously in high-stakes environments.

The image of human test pilots as maverick adventurers, popularised by films like “Top Gun: Maverick,” is far removed from the modern reality, especially in the context of DARPA’s rigorous standards. The agency’s objective is to create a machine-learning agent that can autonomously fly an aircraft without violating training rules. Neural networks, though powerful, are notorious for finding and exploiting unforeseen loopholes, making the stakes of controlling a real jet fighter extremely high. The AI’s success must therefore be measured not just by its performance, but by its adherence to safety and procedural integrity.

The jet fighter at the centre of this initiative is the X-62A Variable Stability In-Flight Simulator Test Aircraft, or VISTA. Originally an F-16D (Block 30) two-seater, the aircraft has spent over three decades at the US Air Force Test Pilot School at Edwards AFB, honing the skills of nearly a thousand students and staff members by simulating the flight characteristics of various aircraft. Its long history and adaptability made it an ideal candidate for the ACE program, leading to its redesignation as the X-62A in 2021.

Initial test flights under AI control began in December 2022, with human pilots on board to monitor and, if necessary, override the AI’s decisions. These early tests involved simulated adversaries, allowing the AI to refine its capabilities in a controlled environment. By September 2023, the program had logged 21 test flights, including the first-ever AI versus human aerial engagement within visual range, pitting the AI-controlled X-62A against a human-piloted F-16. This phase of the program saw over 100,000 lines of flight-critical software changes, reflecting an unprecedented pace of development.

William Gray, chief test pilot of VISTA and the USAF Test Pilot School, emphasised that the program’s goal extends beyond perfecting dogfighting techniques. While dogfighting was a critical challenge to solve, the broader aim is to apply the lessons learned to a wide array of tasks that autonomous systems can perform. This holistic approach ensures that the technology developed is versatile and robust, capable of adapting to various scenarios and challenges in the field of aerospace.

Simultaneously, another groundbreaking achievement in AI has emerged from the world of virtual racing. Sony AI has developed an AI capable of racing at a world-class level in Gran Turismo, a popular racing simulation game. Unlike previous efforts that focused on driving fast, this AI has mastered the art of racing, including tactics, strategy, and racing etiquette. This feat was demonstrated when the AI, named GT Sophy, outperformed some of the world’s best Gran Turismo players in head-to-head competition, as documented in a recent Nature paper.

Racing, much like aerial combat, demands more than raw speed. Effective racing involves real-time vehicle control, strategic overtaking, and maintaining optimal racing lines—all while considering the actions and positions of opponents. GT Sophy was trained using deep reinforcement learning, which involved running numerous scenarios on different car and track combinations in Gran Turismo: Sport. These scenarios varied in track positions, starting speeds, and opponent skill levels, ensuring the AI encountered a wide range of situations.

The AI controlled steering, braking, and acceleration, interacting with the game at a frequency comparable to human players. It was rewarded for progressing along the track efficiently and penalised for errors like corner-cutting and collisions. Within a few hours of training, GT Sophy surpassed about 95 percent of human players, and after 45,000 driving hours over 10 days, it achieved superhuman performance in time trials across multiple tracks.

GT Sophy’s prowess was tested against top-level Gran Turismo players, including Emily Jones, Valerio Gallo, and Igor Fraga. Despite having access to a ghost of the AI’s lap, none of these elite players could best GT Sophy’s times initially. The AI’s ability to discover optimal racing lines and maintain higher corner speeds provided insights that even seasoned racers found valuable. For instance, Jones noted that observing GT Sophy’s lines taught her new strategies for cornering and prioritising exits, while Fraga was impressed by the AI’s ability to maintain speed through corners.

Head-to-head races further demonstrated GT Sophy’s capabilities. In a series of races held in Tokyo, four instances of GT Sophy competed against four top Japanese GT players. Although the humans initially won, a subsequent rematch saw the AI team triumph decisively. The deep reinforcement learning approach enabled GT Sophy to adapt and learn various racing manoeuvres, such as slipstreaming, blocking, and executing different overtakes.

Despite its impressive performance in virtual racing, GT Sophy is not yet ready for deployment in real-world racing or as non-playable characters (NPCs) in Gran Turismo 7. The complexity and variability of real-world environments pose challenges beyond the current capabilities of AI trained in simulations. Nonetheless, the progress made by GT Sophy represents a significant step forward in developing AI that can handle complex, dynamic tasks in both virtual and real-world settings.

These advancements in AI, from DARPA’s autonomous jet fighter to Sony AI’s world-class racing agent, highlight the transformative potential of machine learning in diverse fields. As AI continues to evolve, its applications will likely expand, offering new possibilities and challenges across various domains. The success of these programs underscores the importance of rigorous testing, adaptability, and the pursuit of safe, reliable AI systems capable of performing complex tasks alongside or even surpassing human capabilities.

for all my daily news and tips on AI, Emerging technologies at the intersection of humans, just sign up for my FREE newsletter at www.robotpigeon.be