Probabilistic and optimal Human Navigational Intent Prediction

The ability to predict human behaviour and choose the right action for a robot is crucial to safe and natural operation in an environment shared with humans. Assuming a probabilistic model for human navigational decision-making, in this thesis, we aim to tackle the problem of Human Navigational Intent Prediction. To this end, we propose a probabilistic framework for fast and accurate estimation of the probability distribution over future human states given the previous state. We introduce three internal parameters to our model: the human goal (g), the optimality of human actions (β), and farsightedness (γ). Our framework maintains and updates a belief over the parameters by observing human actions online. Also, in contrast to existing methods, we consider a 4D state for human dynamics. We overcome the challenges introduced by using a more complex model by precomputation of a Time-To-Reach based value function and exploiting particle filter sampling for human states and the goal distribution. We evaluated our method using synthetically generated and real-world data and have shown that our method outperforms the baseline in longer prediction horizons.