Safety, Challenges, and Performance of Motion Planners in Dynamic Environments

Abstract

Providing safety guarantees for autonomous vehicle navigation is an ultimate goal for motion planning in dynamic environments. However, due to factors such as robot and obstacle dynamics, e.g., speed and nonlinearity, obstacle motion uncertainties, and a large number of moving obstacles, identifying complete motion planning solutions with collision-free safety guarantees is practically unachievable. Since complete motion planning solutions are intractable, it is critical to explore the factors that impact planning success. One such factor is the planning environment, e.g., obstacle speed, obstacle motion uncertainty, and number of obstacles. In this paper, we explore the impact of the environmental parameters on the performance of a set of thirteen planning algorithms for navigating in dynamic environments. We aim to answer: (1) How do these algorithms perform relative to each other under increasingly more challenging environments? (2) What factors in an environment make planning in dynamic environments challenging? We classify and compare the algorithms in two planning environments with varying types and magnitudes of environmental challenges. Results show that state of the art planning algorithms were unable to consistently identify collision-free paths even in simple geometric planning problems with moving obstacles with stochastic dynamics. Results also demonstrate that given accurate obstacle predictions, planning algorithms that work in state-time space can typically generate real-time solutions in a limited planning horizon with higher success rates than other methods. In addition, in the presence of obstacle motion uncertainty, accepting paths with non-zero collision probability may lead to higher success rates.