Military drone operations since the 1980s generated hard lessons about remote supervision that the self-driving industry is now relearning through accidents and frozen vehicles. Missy Cummings, who studied UAV remote control systems as a Navy fighter pilot and researcher, identifies five failure modes that plagued early drone programs and now affect autonomous cars: latency, poor interface design, operator workload swings, inadequate training, and weak contingency planning.

Latency is the critical factor. Early Air Force attempts to remotely take off and land drones in the Middle East from Las Vegas produced accident rates 16 times higher than piloted fighter jets due to two-second communication delays. The military switched to automated takeoffs and local line-of-sight control. Self-driving companies rely on cellphone networks and now route supervision through operators in the Philippines, increasing delays further. When a Waymo operator instructed a car to turn left on what appeared to be a yellow light in the video feed, network lag meant the real-world light had already turned red.

Interface design drove between 20% and 100% of human-error UAV crashes across Army and Air Force platforms through 2004. In one case, button placement on controllers made it easy to accidentally shut down an engine instead of firing a missile. Commercial autonomous shuttles now use off-the-shelf gaming controllers never designed for vehicle control, leading to mode confusion crashes. The 2025 San Francisco power outage exposed contingency planning failures when frozen Waymo vehicles blocked first responders — vehicles designed to perform "minimum-risk maneuvers" failed to pull aside.

Operator workload in drone missions swings between crushing boredom during long surveillance periods and overwhelming intensity during strikes. Self-driving car supervisors face the same pattern. The military has struggled for years to have one person supervise multiple drones due to cognitive switching costs and workload spikes. If every autonomous car requires dedicated human attention, the business model collapses. Yet no public training standards exist for remote operators, and no regulations govern their qualifications — unlike aviation dispatchers performing similar roles under FAA oversight.

The research will be presented at the 2026 IEEE International Conference on Human-Machine Systems in Singapore in July. The fundamental tension remains: as long as AI cannot handle uncertainty, remote human supervision is mission-critical, not a support feature. Commercial deployments are treating it as the latter.