Safety‐critical control for robotic systems with uncertain model via control barrier function

Abstract

Usually, it is difficult to build accurate dynamic models for real robots, which makes safety-critical control a challenge. In this regard, this article proposes a double-level framework to design safety-critical controller for robotic systems with uncertain dynamics. The high level planner plans a safe trajectory for low level tracker based on the control barrier function (CBF). First, the high level planning is done independently of the dynamic model by quadratic programs subject to CBF constraint. Afterward, a novel method is proposed to learn the uncertainty of drift term and input gain in nonlinear affine-control system by a data-driven Gaussian process (GP) approach, in which the learning result of uncertainty in input gain is associated with CBF. Then, a Gaussian processes-based control barrier function (GP-CBF) is designed to guarantee the tracking safety with a lower bound on the probability for the low level tracker. Finally, the effectiveness of the proposed framework is verified by the numerical simulation of UR3 robot.

Off‐policy model‐based end‐to‐end safe reinforcement learning

“…This task poses significant challenges, particularly since system dynamics exhibit uncertainty and are subject to changes over time. 3,4 Therefore, a recent focus has been on developing learning-based controllers that can balance the trade-offs between safety and performance whilst considering uncertainties of system dynamics. 5 Reinforcement learning 6 (RL) has emerged as a powerful machine learning tool for designing optimal controllers for uncertain systems by iteratively interacting with the environment.…”

“…There has been a growing interest in the development of optimal controllers that achieve predefined performance while satisfying safety constraints. This task poses significant challenges, particularly since system dynamics exhibit uncertainty and are subject to changes over time 3,4 . Therefore, a recent focus has been on developing learning‐based controllers that can balance the trade‐offs between safety and performance whilst considering uncertainties of system dynamics 5 …”