Intended Human Arm Movement Direction Prediction using Eye Tracking

Artikel i vetenskaplig tidskrift, 2024

Collaborative robots are becoming increasingly popular in industries, providing flexibility and increased productivity for complex tasks. However, the robots are still not interactive enough since they cannot yet interpret humans and adapt to their behaviour, mainly due to limited sensory input. Prediction of human movement intentions could be one way to improve these robots. This paper presents a system that uses a recurrent neural network to predict the intended human arm movement direction, solely based on eye gaze, utilizing the notion of uncertainty to determine whether to trust a prediction or not. The network was trained with eye tracking data gathered using a virtual reality environment. The presented deep learning solution makes predictions on continuously incoming data and reaches an accuracy of 70.7%, for predictions with high certainty, and correctly classifies 67.89% of the movements at least once. The movements are, in 99% of the cases, correctly predicted the first time, before the hand reaches the target and more than 24% ahead of time in 75% of the cases. This means that a robot could receive warnings regarding in which direction an operator is likely to move and adjust its behaviour accordingly.