Architecture-independent analysis of task conflicts in data-driven end-to-end controllers for lower-limb wearable robots

End-to-end control strategies for lower-limb wearable robots map sensor data directly to joint moment predictions, generalizing assistance across locomotor tasks without explicitly classifying them. These control models are iteratively trained/validated with varied sensor inputs and time windows, which is a costly process that conflates the performance of the model architecture with fundamen- tal conflicts in the data. Here we present an architecture-independent framework that quantifies input-output conflicts across tasks using multivariate Gaussian models of phase-dependent biomechanical data. Conflict heatmaps reveal gait phases where similar sensor inputs demand contradictory torque outputs between tasks. Analysis of the hip, knee, and ankle shows the relative efficacy of different input sources and time windows at reducing conflicts. We also find an empirical correlation between input-output conflict and model error for example archi- tectures. Supported by an online tool, this framework enables principled sensor selection and conflict analysis to advance the development of versatile prostheses and exoskeletons.