Human behavioral metrics of a predictive model emerging during robot assisted following without visual feedback

Robot-assisted guiding is gaining increased interest due to many applications involving moving in the noisy and low visibility environments. In such cases, haptic feedback is the most effective medium to communicate. In this letter, we focus on perturbation-based haptic feedback due to applications like guide dogs for visually impaired people and potential robotic counterparts providing haptic feedback via reins to assist indoor fire fighting. Since proprioceptive sensors like spindles and tendons are part of the muscles involved in the perturbation, haptic perception becomes a coupled phenomenon with spontaneous reflex muscle activity. The nature of this interplay and how the model-based sensory-motor integration evolves during haptic-based guiding is not well understood yet. We asked human followers to hold the handle of a hard rein attached to a one-DoF robotic arm that gave perturbations to the hand to correct an angle error of the follower. We found that followers start with a second-order reactive autoregressive following model and changes it to a predictive model with training. The reduction in cocontraction of muscles and leftward/rightward asymmetry of a set of followers behavioral metrics show that the model-based prediction accounts for the internal coupling between proprioception and muscle activity during perturbation responses