By: Dawn Zoldi
At the 2026 GEOINT Symposium, a panel of defense technologists, geospatial innovators, and autonomy engineers discussed how the navigation assumptions built into most autonomous systems are already broken. Fixing them requires rethinking what trust means at the machine level.
The Big Navigation Lie
GPS denial is no longer the worst case. The threat landscape has changed. GPS jamming and spoofing aren’t new, but they are newly prolific and devastating for low-cost uncrewed systems that lack anti-jam capabilities. The distinction matters. A system that knows it is lost can fail gracefully. A system that thinks it knows where it is, but doesn’t, might continue executing a mission with lethal confidence in a lie.
Damien Tyrrell, Regional Director at Shield AI and the panel’s voice from the operational front lines in Ukraine, noted that the real danger is believing a corrupted signal. “The worst case scenario is not the lack of a GPS signal,” Tyrrell said. “The worst case scenario is spoofing and intermittent comms and incorrect data, and that the systems actually rely on or believe that data.”
“We need alternative navigation because these attacks are breaking the assumptions for the development and operation of these systems,” noted Justin Klotz, Ph.D., Autonomy Department Manager at Leidos and principal investigator for DARPA’s Rapid Experimentation for Autonomous Capabilities (REAC) program.
Defining Trust
The panel’s central theme revolved around trust. Nick Bousquet, Vice President of Strategy at GRVTY and a former Navy intelligence officer, surfaced an uncomfortable asymmetry that the entire autonomous systems industry faces. Humans tolerate a higher error rate from other humans than they do from machines. “We are more accepting of a higher error rate from a human and we are terrified of a lower error rate return from something that is an AI or an autonomous system,” he said.
Moderator Michael Robbins, President and CEO of AUVSI, made the same point in civilian terms. When a Waymo vehicle struck a child while likely preventing a fatality, it dominated headlines for days. Yet roughly 40,000 Americans die on roads annually, largely without remark.
For Shield AI’s teams working alongside the Ukrainian military, the operational takeaway as to what to trust in the field has been continuity of behavior. “We’re looking for continuity of behavior as a trust factor for the operators more than absolute accuracy,” Tyrrell said.
Bousquet took the concept further. He described what amounts to an almost zero-trust architecture for any sensor data an autonomous system ingests. “Not any single source can actually be trusted individually,” he said. “What are your fallback measures? And making sure your end users are accessing that in real time is what’s going to establish trust and ultimately let you make operational decisions at the edge.”
State Estimation, The Backbone of Resilience
The lesson from Ukraine and Gaza is that environments change violently and immediately. A visual navigation system cuing off a building that no longer exists is navigating in fiction.
For Tyrrell and Shield AI, the technical answer to GPS denial converges on state estimation. The question isn’t whether incoming data is perfect. The question is whether the system can evaluate the quality of what it’s receiving in real time and act accordingly.
“It would be nice if it was GPS off and then you could just go about your job, but it’s not. It’s on, it’s off. Is it real, is it not? And then you have to correlate across that,” Tyrrell said. “What we’re really talking about is state estimation. Does your state estimation have a wide tolerance for uncertainty?”
Klotz broke down what a modern alternative navigation stack actually looks like in practice. At Leidos, the toolkit includes vision-based navigation, magnetic anomaly navigation, gravimetric anomaly navigation, synthetic aperture radar (SAR), radar-interferometric SAR and signals of opportunity. There is no single dominant solution. “There’s still unfortunately no one size fits all,” Klotz said. “Sometimes it’s all à la carte.”
For low-cost platforms, vision-based navigation often delivers strong performance over land. Signals of opportunity fill the gap over water. Magnetic anomaly navigation, still maturing, holds significant promise, but requires what Klotz called “pristine magnetic data maps” collected in situ. Defense Innovation Unit’s (DIU) quantum sensors transition program is one effort currently advancing that effort, Klotz said.
Hugh Hayden, Head of Go-To-Market at Niantic Spatial, zeroed in on the freshness problem. Niantic’s large geospatial model (LGM) fuses data from ground-based robots, drones and space-based sensors into a continuously updated spatial intelligence layer. “How has that city block changed that morning? Well, look back to the last robot’s data that rolled through that area,” Hayden said. The company’s work with Coco Robotics, whose delivery robots were getting lost in the urban canyons of Los Angeles, turned out to be a direct analog for contested navigation in contested environments. The fundamental questions the system had to answer, Hayden said, were identical: “Where am I? How am I oriented? What am I looking at? How does that thing impact me and how does that inform my next decision?”
Regional Collapse, The Problem Scaled
Robbins posed a scenario that raised the stakes further: what happens when GNSS doesn’t degrade locally, but becomes unreliable at a regional scale, as has occurred in and around the Strait of Hormuz, with no warning whatsoever?
Tyrrell’s answer distinguished between two different categories of failure. “In localized environments when you lose GPS, that’s a navigation issue,” he said. “When you lose it over large regional areas, it actually becomes a mission issue.” At that scale, the concern shifts from flight autonomy, whether the platform knows where it is, to mission autonomy, the layer responsible for real-time decision-making in the field. A system operating across wide areas needs state estimation tolerant of accumulated drift over extended time horizons, not just momentary signal interruptions.
Klotz pointed to the Drone Dominance Program, a Department of Defense effort administered largely through DIU, as a major forcing function. The program purchased 30,000 drones in early 2026 alone, with another 50,000 planned before year’s end and hundreds of thousands projected for 2027. Subsequent “Gauntlets” in the program will require platforms to overcome loss of RF communications, then demonstrate onboard mission autonomy, and eventually operate in swarm configurations. The requirements for each phase are not traditional program-of-record requirements, they function as competition thresholds. (See prior AG coverage of the Drone Dominance Program).
Resilience as Architecture
As platforms begin to operate at scale in multi-agent environments, resilience needs to become a property of the architecture as a whole. “When you actually then start talking about multi-agent swarming… it actually becomes extremely challenging. It’s exponentially challenging,” Tyrrell said. “Rather than discrete resilient systems, we actually start talking about resiliency as an architecture across individual units and multi-agent and then swarming.”
Hayden reinforced this with a call for data interoperability, not just at the hardware or operating system level, but at the foundational data layer. Stovepiped data collection platforms produce stovepiped situational awareness. A unified, queryable geospatial model accessible across autonomous systems actually enables fleet-scale coordination in a degraded environment.
The Testing Gap and the “Last 10 Percent”
No discussion of resilient autonomy would be complete without confronting the gap between what’s needed and what’s currently permitted in a testing environment. Running meaningful electronic warfare (EW) testing in the United States is, in Robbins’ words, “basically impossible.” AUVSI is currently pushing language in the 2026 National Defense Authorization Act to require closer coordination between the Department of Defense, the FAA and the FCC to enable more realistic domestic testing.
Ukraine has functioned, in practice, as the most rigorous test environment available. The companies that stayed there, iterated and built ground teams there are now the ones whose systems work in contested conditions. “The most ideal testing environment is real life,” Tyrrell said flatly.
Klotz recommended to simulate exhaustively, then test at the most challenging available range with all RF denied. He also said, critically, to bring operators and warfighters into R&D as early as possible. They know the threats and where similar systems have failed.
“Know that you’re actually only making it to, let’s say, 90%,” he said. “That last 10% is the unknown variable that your system needs to be able to adapt on the fly. That’s where you make the biggest difference,” Tyrrell suggested.
In short, fully autonomous systems can operate with confidence in GPS-denied environments. Shield AI is proving that today in active conflict zones. But getting there demands abandoning the fiction of absolute positional certainty in favor of something more durable: a system that knows what it knows, knows what it doesn’t and keeps flying (swimming or moving) anyway.