Verifying Learning-Based Robotic Navigation Systems

Abstract Deep reinforcement learning (DRL) has become a dominant deep-learning paradigm for tasks where complex policies are learned within reactive systems. Unfortunately, these known to be susceptible bugs. Despite significant progress in DNN verification, there been little work demonstrating the use of modern verification tools on real-world, DRL-controlled In this case study, we attempt begin bridging gap, and focus important task mapless robotic navigation — classic robotics problem, which robot, usually controlled by DRL agent, needs efficiently safely navigate through an unknown arena towards target. We demonstrate how engines can used effective model selection , i.e., selecting best available policy robot question from pool candidate policies. Specifically, detect rule out that may suboptimal behavior, such as collisions infinite loops. also apply identify models with overly conservative thus allowing users choose superior policies, might better at finding shorter paths To validate our work, conducted extensive experiments actual confirmed detected method were indeed flawed. superiority verification-driven approach over state-of-the-art, gradient attacks. Our is first establish usefulness identifying filtering real-world robots, believe methods presented here applicable wide range systems incorporate deep-learning-based agents.