Content area
Abstract
Limb dysfunction and, in some cases, even amputation is a reality that several people worldwide have to face on a daily basis. Such limitations are frequently associated with traumatic causes, strokes, or cardiovascular diseases. With technological advances, these adversities can be softened with prosthetics or exoskeletons. However, access to these solutions often requires an expensive investment and is not readily available to everyone. Even though one can opt for ad-hoc low-cost alternatives, developing a functional and adaptable prototype is not a trivial task.
The application of this research is to implement a Deep Reinforcement Learning model capable of moving a virtual arm in response to a person’s brain activity. It also aims at exploring and studying the advantages and limitations of using these approaches to solve control problems in prosthetic robotics. With this in mind, the implementation structure can be divided into three main components. Firstly, processing the data that consists of EMG signals from the bicep and tricep. Afterwards, the development and analysis of a supervised learning approach followed by different Reinforcement Learning scenarios. Finally, the work also addresses the generalisation problem by testing a trained model’s reaction when facing a new subject’s EMG signals.
The methodology used different metrics to evaluate the models’ performance to provide a more comprehensive understanding of the system’s behaviour. This assessment takes into account the arm movement’s delay to direction changes and precision. Multiple graphs were made to compare the real and the predicted angles visually. According to this analysis, the algorithm that offered the best results was a Deep Deterministic Policy Gradient version optimised for sample efficiency. It showed great potential to be used in real-life scenarios. However, when approaching the generalisation, this same algorithm showed unsatisfactory results, suggesting that further research may be required.
