DarkstarDarkstarBecome a design partner
Ground autonomy testing

Testing ground robots and off road autonomy

Off road autonomy has to handle terrain, soil and vegetation no map fully captures, and public test results show the technology is less mature than many expected. This page collects what official sources say about ground autonomy programs and why simulation has struggled to predict real performance.

Public sources, checked October 14, 2026.

What the public record says

DARPA's RACER program develops autonomy software, not new vehicles or sensors, so unmanned ground vehicles can drive off road at speeds at least on par with a human driver. Source

RACER Phase 1 experiments ran in March 2022 at Fort Irwin, September 2022 at Camp Roberts and March 2023 at Fort Irwin, and Phase 2 began in fall 2023 and added a 12 ton tracked RACER Heavy Platform. Source

A January 2026 DARPA article says RACER is approaching its conclusion after an eighth and final experiment at Fort Irwin. Source

The Army's Product Manager for Robotic Ground Combat Systems issued a request for information for Unmanned Ground Commercial Robotic Vehicles on August 15, 2025. Source

Why testing is hard

  • DARPA says off road conditions are far less predictable than on road, and in the first Phase 2 experiment teams had no prior exposure to the terrain or sensor data and no practice runs. Source
  • DARPA's RACER Sim announcement said virtual models did not reliably predict real off road performance because of complex 3D terrain, varied soils and vegetation, many obstacle types and limited map data. Source
  • A May 2025 Congressional Research Service report says a 2024 Army off road autonomy assessment found the technology far less mature than many expected. Source
  • GAO found in December 2025 that DoD wide test policies only partly reflect leading practices such as involving testers early, iterative testing, and digital twins. Source

Where Darkstar fits

Darkstar is an early stage simulator project for replaying adversarial scenarios against autonomous systems on real terrain, designed for drone and swarm autonomy teams first. It is not a finished product: design partners help decide which scenarios it replays first, so the work fits their own evaluations.

Other segments

Guides

Darkstar by Think Tech Holdings

Find autonomy failures in simulation, not in the field

Darkstar is an early stage simulator project for replaying adversarial scenarios against autonomous systems on real terrain. We are looking for design partners to shape what it tests first.

Become a design partner
DarkstarGuidesPrivacy and text terms