Risk-Aware Navigation on Smooth Approximations of Euclidean Distance Fields Among Dynamic Obstacles

You are here: Homepage > Search

Risk-Aware Navigation on Smooth Approximations of Euclidean Distance Fields Among Dynamic Obstacles

Free access

Barbosa, Fernando S. / Tumova, Jana

Reasoning about probabilistic outcomes on stochastic models is of essence to safety-critical systems. In this paper we focus on risk-aware collision avoidance approaches in workspaces with static and dynamic obstacles. More specifically, for Lagrangian systems operating in workspaces without perfect information about the obstacle-space. In order to avoid collision with static obstacles, we build smooth approximations of the Euclidean distance field, along with its first and second derivatives, using Gaussian Process implicit surfaces. Since the predictive distance returned by such an approximation is a normal distribution, rather than simply using its mean value, we propose a risk-aware Control Barrier function. Risk metrics provide more coherent measures than chance constraint, with the benefit of distinguishing between tail events. We prove that by using the proposed approach, the Lagrangian system is bound to a smaller, but safer (in terms of risk-awareness), subset of the obstacle-free space. Besides that, we also propose a controller for avoiding collisions with ellipsoidal dynamic obstacles. We compose all the controllers together into a nonsmooth barrier function, and design a Quadratic Program-based optimization controller. The proposed approach is a step-forward towards closer integration between mapping algorithms and feedback controllers. Numerical simulations on synthetic environments highlight the capabilities of the approach proposed. ; QC 20220112