A Simple Neural Network for Collision Detection of Collaborative Robots.

Area of Science:

• Robotics
• Artificial Intelligence
• Sensor Technology

Background:

• Increasing automation necessitates robots safe for human interaction, especially in space-constrained or security-limited environments.
• Traditional safety measures for collaborative robots (cobots) often involve expensive force sensors.
• There is a need for cost-effective collision detection solutions for cobots.

Purpose of the Study:

• To present a practical collision detection method using a simple neural architecture as a virtual force and torque sensor.
• To evaluate the effectiveness of different neural network architectures for this application.
• To assess the performance of the developed virtual sensor on a cobot prototype.

Main Methods:

• Implementation of a virtual force and torque sensor using neural networks.
• Comparison of four architectures: Auto-Regressive (AR), Recurrent Neural Network (RNN), Convolutional Long Short-Term Memory (CNN-LSTM), and Mixed Convolutional LSTM (MC-LSTM).
• Analysis of architectures at various input regression levels (motor current, position, speed, control velocity) and testing on the CURA6 cobot prototype.

Main Results:

• The MC-LSTM architecture proved most effective at a regression level of 12 samples (24 Hz), achieving a mean absolute prediction error of approximately 22 Nm.
• External tests with 72 collision signals confirmed the architecture's viability as a collision detector.
• The MC-LSTM achieved an f1 score of 0.85 for collision detection with an optimal threshold.

Conclusions:

• A virtual sensor based on the MC-LSTM neural network architecture can effectively detect collisions in collaborative robots.
• This approach offers a cost-effective alternative to traditional force sensors for cobot safety.
• The developed virtual sensor has potential applications in various human-machine interaction systems requiring collision detection.