Uncertainty-Aware Knowledge Distillation for Collision Identification of Collaborative Robots

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Collaborative robots are increasingly used in industry, necessitating robust safety features.
• Collision identification is crucial for preventing accidents in human-robot interaction.

Purpose of the Study:

• To propose a deep learning method for identifying external collisions in 6-DoF articulated robots.
• To enhance the accuracy of collision identification by addressing uncertainties in deep neural networks.

Main Methods:

• The study expands on the CollisionNet concept to pinpoint external force locations.
• A novel uncertainty-aware knowledge distillation technique is employed.
• Sample-level uncertainties are estimated from a teacher network to guide student network training with focused penalties.

Main Results:

• The proposed deep learning method effectively identifies external collisions in articulated robots.
• Uncertainty-aware knowledge distillation significantly improves the accuracy of collision identification.
• Experimental results validate the method's effectiveness in enhancing robot safety.

Conclusions:

• The developed deep learning method offers a promising solution for external collision identification in collaborative robots.
• Uncertainty-aware knowledge distillation is a key contribution for improving deep neural network performance in this domain.
• This advancement contributes to safer and more reliable human-robot collaboration in industrial settings.