Temporal signatures of prediction error reveal external forces on an embodied robot

Embodied robots must distinguish self-generated sensory consequences from externally caused perturbations. In predictive processing, such deviation is accounted for as prediction error, but it is commonly treated primarily as a signal to be minimised. Here, we test whether the temporal profile of load-related prediction error provides diagnostic information about the type of physical interaction a robot experiences. We trained and evaluated windowed multilayer perceptrons to predict servo current, used as a proxy for joint load, from proprioceptive, inertial, and servo current history. The training data were recorded during random movements of tilt and pan servo motors of a robotic head. The best model was deployed in the robot in real time and evaluated across free movements and external perturbations, including brief pushes, sustained forces, and obstacle interactions. Prediction error and raw current time series were summarised using features capturing magnitude statistics, temporal structure, and signal shape. A random forest classifier was then used to classify the conditions based on the features. Different interaction conditions produced distinct profiles that supported robust perturbation classification with high cross-validated accuracy. Classification performance for prediction error was equal to the classification of raw current. The results suggest that load-related prediction error is not only a residual of imperfect prediction. Its temporal structure can act as an embodied self-monitoring signal for classifying physical interactions.