Robotic Fault Detection using Temporal GANs

Voiceover

Presenter(s)

Shyam Sai

Abstract or Description

Every robot requires fault detection to ensure the robot's safety, especially in environments where the robot cannot be easily accessed by humans. Fault detection has traditionally been done similar to expert systems, though deep learning systems have begun to appear. This project uses a temporal GAN architecture - namely TimeGAN - to perform fault detection on a robot. The TimeGAN architecture is a typical GAN architecture but is better suited to time-series data like robot movement by using recurrent neural networks as components. The Gazebo simulator and open-source Turtlebot model is used to collect training and testing data for this approach. Odometry data is collected both in normal driving conditions and in faulty conditions (broken wheel, obstacle collision). Results show that the TimeGAN generator is successfully able to generate data of a similar distribution to the normal data collected from the robot. Furthermore, the GAN's discriminator is successfully able to differentiate between normal data and faulty data, showing the potential of a temporal GAN architecture's use in robotic fault detection.

Mentor

David Wettergreen