CAMBRIDGE, Mass. -- This week, the Department of the Air Force-Massachusetts Institute of Technology Artificial Intelligence Accelerator will commence the CogPilot Data Challenge 2.0. The challenge invites participants to explore AI-based solutions linking a pilot’s physiology, such as heart rate, eye tracking, and muscle activity, to their behavior and performance of flying tasks varying in difficulty. The AI Accelerator welcomes all participants across the Department of Defense, academia, and industry, and aims to accelerate innovation by engaging the broader AI community in attacking tough DoD technology needs.
Since its inception in 2019, the AIA has released more than 10 challenges from its various research projects. “These challenges have turned into some of our greatest successes at the AI Accelerator,” said Maj. Kyle “Gouge” McAlpin, Performance Prediction and Optimization project liaison. “They have surfaced and fostered organic machine learning talent across the Air and Space Forces, built vibrant communities of cross-disciplinary researchers and operators pushing the state-of-the-art, invited the public to join in solving some of our hardest problems, and given back to the machine learning community by funding and releasing large, machine learning-ready public datasets. The machine learning community has found time and time again that fundamental advances in ML start with strong competition on large, unique, and public datasets.”
The CogPilot Data Challenge 2.0, hosted by AIA’s Performance Prediction and Optimization research team, consists of two tasks. First, participants are challenged to develop a model that predicts the difficulty level of an aircraft landing performed in virtual reality based only on pilot physiology. For the second task, participants predict how well the pilot performed each approach and landing task using only pilot physiology.
To collect the dataset, the team used an immersive, virtual reality simulator to record many different types of physiological measurements while pilots performed approaches and landings to a runway at four varying difficulty levels. The scenario used a simulated T-6A Texan II fixed wing trainer aircraft flown with a stick, throttle, and rudders. The CogPilot Data Challenge 2.0 is a continuation from a previous challenge which hosted over 180 participants. CogPilot 2.0 provides additional data from 20 pilot participants along with new awards.
The Performance Prediction and Optimization research team is exploring how quantitative performance measurements and physiological monitoring can provide a more individualized and objective assessment of pilot training compared to current subjective, coarse measures. Partnering with pilot training units, the team uses AI to incorporate various physiological assessments into a single measure of cognitive workload. The cognitive workload measurement may be used to accelerate debriefings by targeting portions of a flight or simulation which induced high cognitive workload. It may also be used by instructor pilots to tailor their instruction in flight.
The CogPilot Data Challenge 2.0 registration will be open through January and challenge submissions are due February 15th. Team results will be posted March 7th. To learn more and register, visit https://pilotperformance.mit.edu/.