Calspan

From Skepticism to Acceptance, the Auto GCAS Story

Here is a BFO* for your consideration. Pilots like flying airplanes. It’s why they (we) became pilots to begin with. So, when the subject of introducing autonomy and artificial intelligence into airplanes comes up, you’re likely to find the primary customer skeptical at best, hostile to the subject at worst. Developing trust that the autonomy implementation adds value is key to overcoming those objections. One historical example demonstrates the case well; the implementation of Automatic Ground Collision Avoidance System (Auto GCAS) in the United States Air Force.

Between 1992 and 2004, the F-16 community experienced 34 F-16 Controlled Flight Into Terrain (CFIT) mishaps resulting in 24 fatalities. NASA and USAF created the Auto GCAS program in 1997 to address the issue. The new capability was not initially embraced and was even halted in 2003 due to conflicting opinions of the system’s utility. The pilot community raised concerns over false alerts potentially resulting in unnecessarily aborted bombing runs. After more than a decade of testing and updates, it was fielded broadly in 2014, and by 2020 had been credited with at least 12 F-16s and 13 pilot lives saved. Having earned the Air Force’s trust, the system has also been integrated into the F-22 and F-35 platforms, and the US Navy is integrating the capability into the FA-18 E/F and EA-18G.

* Blinding Flash of the Obvious

What Testing Has Taught Us About AI in the Cockpit

This program demonstrates an autonomy implementation process which employed a bespoke testbed aircraft, the Variable In-Flight Stability Test Aircraft (VISTA), recently re-designated the X-62A. Integrating Artificial Intelligence (AI) features into fielded or developmental aircraft will likely follow similar pathways: Initial resistance, laboratory testing, then real-world testing in bespoke surrogate testbeds, limited initial fielding to build trust, then accelerating wide acceptance. However, an autonomy system like Auto GCAS has a delineated input-output loop, whereas non-deterministic AI may respond in surprisingly different ways when stimulated by seemingly identical environmental/situational factors, so the test problem grows to developing confidence in AI’s performance in order to certify and proliferate technology that produces non-deterministic outcomes. A recent example of this was DARPA’s Alpha Dogfight Trials, followed by DARPA Air Combat Evolution (DARPA ACE). In these experiments, offerors submitted AI agents capable of learning the art of fighter maneuvering, or “dogfighting”, to compete against other offeror’s algorithms in a bracket style competition of synthetic engagements. The agents learned to minimize track crossing rate to employ air-to-air gunnery or missiles as rapidly as possible. Unfortunately, that strategy resulted in many head-on collisions between agents that need not consider surviving a crash, doing nothing to earn their human evaluators’ trust. Provided with full situational awareness of their aircraft state and that of their adversary, the test series accepted these artificialities in order to allow the algorithms to learn from each other’s mistakes and successes. At the end, the most successful algorithm was pitted against a human fighter pilot flying a simulator and using virtual reality goggles to view the algorithm’s projected adversary aircraft. The algorithm beat the human pilot in 5 out of 5 engagements, unsurprisingly since it held perfect knowledge of the human’s aircraft state, while the human was deprived of peripheral vision and proprioceptive cueing and having no indicator of the algorithm-pilot’s aircraft energy state. Still, the exercise showed human pilots that an algorithm can learn very rapidly, sowing the seeds of trust that the Services hope will germinate into collaboration between human and synthetic pilots in the Collaborative Combat Aircraft (CCA) now in development. Subsequent efforts using USAF’s X-62A VISTA have demonstrated some promise, enabling Air Force Secretary Frank Kendall to experience dogfighting via AI Pilot first-hand. Similar autonomy testbed aircraft, like Calspan’s Learjetbased Autonomy ExPERTTM**, are being employed to develop autonomous aerial refueling (AAR) capabilities for uncrewed aircraft and AI-piloted CCAs.

“As the algorithms’ performance improves, it can be applied in more complex scenarios with greater trust—precisely how human pilots are certified.”

**Experimentation Platform for Enhanced Research and Testing

The key takeaway from these developments has been process. As President Ronald Reagan famously counseled using a Russian proverb, “trust, but verify”. At least for the near-term, non-deterministic AI needs to demonstrate itself under a variety of real-world scenarios in fault-tolerant systems using run-time assurance (RTA), the ability to selfmonitor and verify or interrupt hazardous algorithm outputs. Mirroring human learning, the algorithms’ exposure to realistic scenarios gives them the opportunity to adapt their behaviors not only from success vs. failure metrics, but from the corrections it garners from surrogate testbeds’ RTA routines. As the algorithms’ performance improves, it can be applied in more complex scenarios with greater trust— precisely how human pilots are certified.

What’s old is new again!