Bayesian Optimization with Embedded Stochastic Functionality for Enhanced Robotic Obstacle Avoidance

Designing an obstacle avoidance algorithm that incorporates the stochastic nature of human-robot-environment interactions is challenging. In high risk activities, such as those found in nuclear environments, a comprehensive approach towards handling uncertainty is essential. In this article, in the context of safe teleoperation of robots, an automated iterative sampling procedure based on Bayesian optimization is proposed, where the robot is trained to predict the behaviour of a human operator. Specifically, a Gaussian process regression model is used to learn an effective representation of a safe stop manoeuvre, required for implementing an obstacle avoidance shared control algorithm. This model is then used to predict the future time duration to execute a safe stop manoeuvre, given the current real-world circumstances. The control algorithm expects this value to be reasonably high; if not, it will gradually reduce the human operator's authority. A distinctive attribute of the proposed method is the use of statistical confidence metrics as tuning parameters, intended to provide a statistical indication of whether or not an obstacle will be avoided. The proof-of-concept experiments were carried out using three robotic platforms suited for use in nuclear robotics, an amphibious SuperDroid HD2 robot equipped with a Velodyne VLP16 (a 3D lidar), an AgileX Scout Mini R&D Pro land robot fitted with a Realsense D435 depth camera, and a Husarion ROSBot 2.0 Pro supplied with an RPLIDAR A3 (a 2D lidar). The test results show that the proposed Bayesian optimization method uses 8 times less data compared to an exhaustive grid approach, and that it provides a robot-agnostic, robust obstacle avoidance.