A REINFORCEMENT LEARNING METHOD FOR DYNAMIC OBSTACLE AVOIDANCE IN ROBOTIC MECHANISMS

摘要

The paper introduces a robotic mechanism for locomotion in unconventional environments such as aerial rigid lines and reticulated structures and presents a method for real-time dynamic obstacle avoidance. The method is biologically inspired and based on perceptual feedback and reinforcement learning control. The proposed collision avoidance approach does not use any formal representation of the existing obstacles and does not need to compute the kinematic equations of the robot. The obstacle avoidance problem is modeled as a multi-objective optimization problem and can be straightforward applied to any articulated mechanism, including conventional manipulator arms.