In this work, we tackle the problem of intent disambiguation using Fisher Information. Fisher information is widely used in parameter estimation and is a measure that quantifies the amount of information a random variable carries regarding an unknown parameter.
The objective is to use Fisher information to characterize the intent disambiguation capabilities of different control modes. In order to evaluate how good is it for the human to continue to operate in a particular mode, we forward project the belief state for a fixed amount of time. The negative entropy of this projected distribution is taken as a measure of disambiguation. The Expected Fisher Information Density over different subspaces/control modes is a direct measure of the disambiguation capabilities.
The Expected Information Density is given by
In the following example, the goals are shown in black and are maximally spread along the Y (green) dimension. Intuitively moving the y-direction will disambiguate the intended goal easily.
The best disambiguating mode is computed for a grid in the entire workspace and it can be seen that the algorithm overwhelmingly picked ‘y’ as the best disambiguating dimension.
Currently, we are performing an extensive simulation-based study for point robots and robotic arms to investigate the impact of the choice of different parameters, inference schemes, disambiguation algorithms etc.