Prediction-Error Negativity in Physical Human-Robot Collaboration
- Publication Type:
- Thesis
- Issue Date:
- 2021
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Robotic systems for physical Human-Robot Collaboration (pHRC) are often controlled using control systems based on the admittance or impedance of the system. The interaction forces exchanged between the robot and the human co-worker during pHRC may affect the human cognitive state. In pHRC systems, the human cognitive state is often neglected. It is hypothesised that admittance dynamics of the robot have an effect on the human co-worker's cognitive state which can be used to estimate the predictability of the robot behaviour, or simply called the robot predictability. By using an electroencephalogram (EEG) device, the brain activity of the human co-worker can be measured. A feature, called Prediction-Error Negativity (PEN), that can be found in the EEG signal and is visible in the Event-Related Potentials (ERP) has the potential to be used to objectively assess the robot predictability. This thesis addresses the following research question: can the human cognitive state be used to assess and improve the robot predictability during physical human-robot collaboration?
Firstly, the relationship between PEN and changes in the robot admittance is investigated. Changes in the robot admittance were the result of the introduction of resistive forces with first-order dynamics. An analysis of the ERP is performed in the time-domain, to determine whether different admittance dynamics result in different PEN amplitudes. It is found that admittance dynamics can modulate PEN and thus robot predictability. Secondly, six different machine learning classifiers are then compared for classification of PEN by using the data sets collected. A two-class classification problem and a three-class classification problem are formulated for the comparative study. A Convolutional Neural Network (CNN) is found to perform best in both the formulated classification problems, when compared to the other classifiers tested. Thirdly, a singularity avoidance strategy is implemented in a practical pHRC robot and is chosen to assess whether PEN can be detected during pHRC in real applications. The relationship between PEN and human preferences is also investigated and confirmed. Finally, a PEN-based closed-loop control is implemented and it is found that this can reduce PEN by automatically tuning parameters in a singularity avoidance strategy.
Please use this identifier to cite or link to this item: