DARPA's RACER robots drive and detect alone, but a person approves every engagement

At a DARPA demonstration on 15 October 2026, four robot vehicles did almost everything for themselves; the one step they did not take was the decision to attack.¹ They drove, searched and reported at St. Vith Range, Fort Knox, Kentucky, in front of two dozen visitors from military and government organisations.¹ All four come from RACER, short for Robotic Autonomy in Complex Environments with Resiliency, a DARPA effort that pulls together three companies' work on vehicle design, algorithms and secure communications.¹ Task & Purpose sums the programme up as an attempt to let ground vehicles work over rough ground with no human at the wheel.²

Everything began with a simulated threat cue drawn from satellite imagery, which an operator used to send the robot team out.¹ The lead vehicle, with a drone shut inside a box on its back, went hunting alone and froze the moment it saw something that might be a target, holding there for orders while the others came up behind it in stages.¹ The strike, when it came, was a simulated rocket, and its target stood on the next hill, some 300 yards away; three of the four vehicles carried drone payloads of their own.¹ Nothing was engaged until the mission commander approved, an Army photo caption notes of the strike element, which paired Anduril one-way attack platforms with Overland AI's ground vehicles.¹

Dr Jason Gregory, a DARPA technical advisor, drew the line plainly: moving, detecting and manoeuvring were left to the machines, engaging was not.¹ For each detection the vehicles pass a person what he called a chip of the scene downrange, and that person's approval comes before any engagement.¹ Task & Purpose framed the same step from the operator's side: the finds were there to be looked over, and an explosive drone would go only if the operator chose to send it.² DARPA tied the design to Pentagon rules; under Defense Department policy, a decision to use lethal force has to have a human overseeing it.²

The request came from U.S. European Command, which wanted robots able to push out ahead of friendly lines and spot threats early, with Fort Knox playing the part of the European theatre.¹ Gregory called the event the first test of the whole chain, from imagery taken in orbit to vehicles on the ground, after earlier rounds of component and integration testing at the companies' own sites.¹ Getting from this demonstration to something usable in combat is left, in DARPA's hopes, to other Pentagon programmes; for now Task & Purpose calls the technology experimental.² In December DARPA plans another test, this time to hear from soldiers about how the vehicles ought to be employed.¹

The rule in the background is Department of Defense Directive 3000.09, revised in January 2023, which requires weapons to let commanders and operators exercise "appropriate levels of human judgment over the use of force".³ Two further demands in the directive, as the Congressional Research Service summarises them, bear directly on design: an interface that trained operators can readily understand at the moment they decide, and training, tactics and doctrine that show them, under realistic conditions, where the autonomy's limits lie.³ CRS also notes that the directive classes systems by what the human does in selecting and engaging targets, categories often described as a human in, on or out of the loop.³

Where the human sits in the loop

Parasuraman, Sheridan and Wickens built their levels-of-automation model around a point that the word "autonomous" hides: one system can be heavily automated at one stage of the work and hardly at all at the next.⁴ The stages they use run from gathering information and making sense of it to choosing a course of action and carrying it out.⁴ Read that way, the Fort Knox vehicles sit high on sensing and flagging, and on carrying out their own movement, while the choice to engage is kept for a person. Their yardstick for such a split is the operator: what it does to human performance comes before questions of machine reliability.⁴ That puts the weight on something the demonstration reporting leaves out — what the operator actually does between a chip arriving and an approval going back.

KEY IDEAThe machine finds and drives; the choice to engage stays with a person.

The chip as a decision display

A chip is a small, cropped summary of something the machine has already judged worth showing. Automation bias, in Parasuraman and Manzey's review, takes two forms that matter here: a person misses what the aid never pointed to, or acts on something it pointed to wrongly.⁵ Training does not remove these errors, they report, and experience is no shield either; experts made them as well as newcomers when the aid was imperfect.⁵ In their account, complacency is above all a problem of a busy operator: it shows up when watching the automation is one task among several.⁵ Army captions describe an operator supervising four vehicles at once; the reporting does not say how the work was divided between that operator and the mission commander who approved the strike.¹ The directive's call for an interface trained operators can readily understand describes an outcome, not a specification; the models above suggest what would need measuring to show it has been met.³

Questions the methods would ask before December

• How much context travels with each chip: the vehicle's confidence, its position, what it saw just before and after, and whether the operator can ask for more without delaying the decision?
• Does the approval step leave enough time and information for a genuine judgment, or does it drift towards confirming whatever the vehicle proposes — the commission error Parasuraman and Manzey describe?⁵
• What happens to things the vehicles did not flag? If the operator's view is built from detections, omission errors may be invisible by design.⁵
• How does approval workload grow as vehicles and payloads are added, and at what point can one supervisor no longer give each detection the human judgment the directive asks for?³
• Will the December soldier sessions measure performance — decision time, false approvals, missed targets — alongside stated preference, given that the levels-of-automation model treats human performance as the first test?⁴