Learning Safe-Critical Control Using Stochastic CBFs in Uncertain Environment

In this paper, we study the problem of learning safe-critical control in uncertain environments using stochastic control barrier functions (CBFs). We propose a novel learning framework based on real-time sensor data to estimate CBFs in uncertain environments, which can achieve the desired goal while considering safety and low uncertainty. First, we obtain the data samples via sensors in real time and generate the train data set. Secondly, we propose two constraints on the mean function and the covariance function in the process of using the Gaussian regression model to learn the unknown environment, which ensure the safety of the robot and reduce the uncertainty caused by measurement noise and observation of environmental limits. Finally, we define two ZCBFs to satisfy the above constraints and implement quadratic programming control, where the mean is constrained to the robot’s safety and the covariance is constrained to a small uncertainty. In addition, to handle the target navigation for the complex environment, we also design a planning approach to generate the next waypoint for nominal control, choosing the position closest to the target in the circular region with the radius min-mean centered on the self-centered region as the next waypoint. The effectiveness of the proposed methods in dealing with safety and low uncertainty in robotic work in uncertain environments is verified on two different stochastic cases.