“Provably Safe” in the Wild: Testing Control Barrier Functions on a Vision-Based Quadrotor in an Outdoor Environment

Abstract

As quadrotors are increasingly deployed into highly dynamic, safety-critical applications, we must develop safe control methods that work under real-world, uncertain conditions. However, most control methods that guarantee safety are tested either in simulation or in heavily-controlled lab settings, with limited experimentation done in the real world. We wish to identify key roadblocks to deploying safe control methods in the wild. We implement a safe controller using exponential control barrier functions on a quadrotor system with onboard computing and visual state estimation. We perform a series of 16 field tests, of which 4 fail to maintain safety. Key reasons for failure include sudden state estimation error, unaccounted for exogenous disturbances, and a response delay to commanded acceleration. Finally, we propose future work to bring guaranteed-safe methods into the real world.