Toward Optimal Parallel Architectures for Transtibial Prostheses: Kinematic Evaluation of 3-SPS-1-S, 3-UPU-1-S, and 3-UCU-1-S Mechanisms Using ANN-Enhanced Methods

This work compares and analyzes three parallel mechanisms aimed at transtibial lower-limb prosthetics: 3-SPS-1-S, 3-UPU-1-S, and 3-UCU-1-S. It addresses the limitations of current serial-based prostheses, such as discomfort and restricted mobility, and explores the underutilization of parallel architectures in this field. The study focuses on forward and inverse kinematics, workspace, and singularities. Forward kinematics are solved using a hybrid method that combines the Newton-Raphson numerical approach with initial estimates generated by an artificial neural network (ANN). In contrast, inverse kinematics are approached geometrically using actuator lengths and previous positions as inputs. The selection of the 'optimal' mechanism is based on two main criteria: (i) how well the workspace matches the physiological range of the human ankle and (ii) the number and distribution of singularities in that region. Additionally, the ANN training process was adapted to each mechanism's configuration: for the 3-UPU-1-S and 3-UCU-1-S architectures, three hidden layers with 35 neurons each were used, whereas the 3-SPS-1-S mechanism required only two layers with 60 neurons each. These tailored configurations led to improved network performance for each specific mechanism. Ultimately, the mechanism with the fewest singularities and most excellent orientational coverage was identified as the most promising candidate for future prosthetic applications.