Intent Inference using Dynamic Field Theory

Any assistive system needs to have a good idea of what the human’s underlying intent in order to provide appropriate kinds of assistance quickly and accurately. A standard approach is to use the Bayesian framework in which the belief over the goals is iteratively updated using Bayes Theorem. In this work, we propose an alternate and novel intent inference scheme inspired by Dynamic Field Theory in which the time evolution of the probability distribution over goals denoted as p(t) is specified as a dynamical system with constraints.

Dynamic neural fields originally were conceived to explain cortical population neuronal dynamics, based on the hypothesis that the excitatory and inhibitory neural interactions between local neuronal pools form the basis of cortical information processing.

Recurrent interactions between the state variables (individual goal probabilities), robustness to noise and inherent memory make dynamic neural fields an ideal candidate for an intent inference engine.

The full specification of the neural field is given by