Deep Edge-AI for Prosthetic Control: Feasibility of ISPU-Based Solutions for a Robotic Extra Limb

Robotic extra-limbs, such as the Sixth Finger, offer significant potential in enhancing human capabilities. These devices require precise real-time control, making sensing crucial for achieving this goal. In this study, the integration of Inertial Measurement Units (IMUs) and Deep Edge AI in controlling a prosthetic finger was investigated. The proposed approach employs a custom-designed neural network trained on sensor data, providing intuitive control for prosthetic devices. The neural network was trained on a dataset consisting of 6-axis IMU acquisitions, including movement data for opening and closing the prosthetic finger. Through advanced preprocessing techniques, 84 features were extracted, and feature selection using LASSO regression reduced the dataset to 36 relevant features. The network achieved a classification accuracy of 95.10%. By deploying the network directly on the IMU's embedded Intelligent Sensor Processing Unit (ISPU), inference is conducted on the device, eliminating the need for external computing resources. This integration significantly reduces latency and power consumption, facilitating real-time, energy-efficient prosthetic control. Initial deployment tests on the LSM6DSO16IS sensor demonstrate the viability of this Deep Edge-AI approach, with the network requiring minimal memory (400 Bytes of RAM) and offering rapid inference times (21.43 ms) and introducing a modest power overhead of 12.5%.