A Bayesian Deep Neural Network for Safe Visual Servoing in Human-Robot Interaction.

Area of Science:

• Robotics
• Artificial Intelligence
• Human-Robot Interaction

Background:

• Safety is critical in human-robot interaction (HRI).
• Existing collision avoidance methods often lack machine learning integration or fail to address data uncertainty and task efficiency.
• Few studies explore data-based repulsive action selection for collision avoidance.

Purpose of the Study:

• To develop a system for collision avoidance with human hands during image-based visual servoing (IBVS) tasks.
• To utilize a Bayesian deep neural network (DNN) for learning repulsive actions and handling uncertainty.
• To improve the safety and efficiency of HRI tasks through robust collision avoidance.

Main Methods:

• A deep neural network (DNN) was transformed into a Bayesian DNN using Monte Carlo dropout (MC dropout).
• The Bayesian DNN was trained to learn repulsive positions for hand avoidance during an IBVS task.
• The system was implemented and tested on a UR10 robot for robustness and convergence speed.

Main Results:

• The Bayesian DNN demonstrated adequate accuracy and generalization on unseen data, with high predictive interval coverage probability (PICP) values (0.84 for x, 0.94 for y, 0.95 for z).
• The system effectively avoided collisions with human hands and prevented the robot from reaching unsafe poses.
• The Bayesian DNN enabled faster IBVS convergence compared to traditional repulsive poses and showed greater robustness than a standard DNN in unseen environments.

Conclusions:

• Bayesian DNNs offer a robust approach to collision avoidance in HRI, particularly for uncertain environments.
• The proposed system enhances safety by preventing collisions and improves task efficiency by allowing faster convergence.
• This method provides a significant advancement in integrating machine learning for safe and efficient human-robot collaboration.