Marines Fooled A DARPA Robot With A Hilarious Metal Gear Solid Trick

AI has been in the headlines quite a bit lately, from ChatGPT and article writing to assisting in identifying odors with biosensors. But, of course, if the public is fascinated with it, so too is the US Government for various reasons. Per excerpts from Paul Scharre’s unreleased “Four Battlegrounds: Power in the Age of Artificial Intelligence,” the US government and the Defense Advanced Research Projects Agency (DARPA) are using AI to detect human movement in an area, but that isn’t quite going to plan.

AI, and more broadly computer vision, has been around for quite some time for detecting motion and identifying people in photos and video. However, there have been flaws and errors, including misidentification and means to subvert the systems. These sorts of issues cannot be present in a system deployed for the military, so DARPA likely has a rigorous testing regimen, including human trials with AI. But, when testing a system for motion detection, you would never guess what it took to break the algorithm.


Eight crafty U.S. Marines were easily able to subvert the AI algorithm designed and trained for detecting humans walking per Scharre’s “Four Battlegrounds.” On Twitter, The Economist’s defense editor Shashank Joshi posted an excerpt from the book outlining this test with anecdotes from an unknown source named “Phil.” Phil explained that the Marines defeated the system, placed in the middle of a roundabout, in several clever ways, such as the first two who “somersaulted for 300 meters.” Another two “hid under a cardboard box,” and it was reported that “you could hear them giggling the whole time,” which isn’t very Snake-like, but funny nonetheless.


Phil’s favorite, though, is a Marine who “field stripped a fir tree and walked like a fir tree,” and “you [could] see his smile, and that’s about all you [could] see.” What all of this came down to a twist on the basic principles of camouflage, breaking up the outline and contours of your body in an effort to blend in. In this case, video game principles were applied such that the Marines got to be a Dark Souls player, Solid Snake in a box, or just a whole tree like PropHunt for Garry's Mod. The AI had not trained for these scenarios and was simply unable to adapt.

Ultimately, AI trained on one specific task is not very flexible, and we are not likely see this be effectively used in day-to-day life or combat. There are just too many variables with humans that make it difficult, even when something that would not fool a normal human, like a person in a box, comes up. We would bet a decent Halloween costume and a funny gait would break the system as well, so there is a ways to go for AI. Either way, let us know what you think of this amusing test in the comments below. If you want to read this story for yourself, you can preorder “Four Battlegrounds: Power in the Age of Artificial Intelligence” on Amazon.