Utilization of a nonlinear autoregressive neural network with exogenous inputs for foot trajectory generation for upslope terrain.

The rhythm in which robotic prostheses have evolved over the past two decades indicates that they have a promising potential to replace passive prostheses in the foreseeable future. Powered lower limb prostheses, unlike the passive ones, can provide net positive work during the late stance, thus reducing the metabolic cost. Moreover, robotic prostheses can provide different characteristics in compliance with the walking terrains. The conventional three-level controllers of robotic lower limb prostheses are known for their intermittent and nondynamic properties, which limit prosthesis functioning to certain predefined tasks and require intense calibration for the controller's parameters for each user. In this study, we report the capability of a nonlinear autoregressive recurrent neural network with exogenous inputs of generating the foot patterns for upslope terrain, that is, 10° incline, which is unknown for the neural network (not included in the training dataset). A new set of evaluation data from six able-bodied subjects (average age 31 and average weight 70.2 kg) was used to assess the network performance. The results show that the NARX was able to estimate foot trajectory for the new terrain without any further training, with an average RMSE of 2.953° for all six subjects.