Deep Networks and Sensor Fusion for Personal Care Robot Tasks—A Review

摘要

Welfare support systems involving robots have been actively researched during the last two decades. Early attempts have been largely based on classical approaches to process sensor data. The thrust toward high accuracy and reliability along with the need to handle sophisticated tasks has recently driven the exploration of methods that handle data from one or more sensors via deep networks. However, the introduction of deep networks leads typically to increased processing times and enhanced system complexity. This article summarizes the classical approaches and then discusses in detail the penetration of contemporary approaches in artificial intelligence for sensory data processing in the context of robot-assisted personal care. Observations about the spatial and temporal nature of certain tasks are made to determine appropriate network architectures for processing sensor data. In other words, we present one classification at the task level and also study the role of contemporary techniques based on attention mechanisms (and specifically transformers). We, then, present another classification based on the approach for fusing sensor data and study this with reference to the introduction of deep learning methods including generative models. This article also attempts to answer the question: are there personal care robotic tasks that can benefit from the fusion of sensor data along with fusion of contemporary network architectures? Potential areas for further research are also indicated.