Karthik Raju

and 1 more

AbstractIn this paper, we hope to suggest and experiment with novel ways to predict the movement of individuals in a shared space. We use an LSTM as the primary framework behind the prediction model and test alternative ways to modify the predictions, namely a Social Pooling layer and a Dynamic Bayesian Network.IntroductionIn recent years, autonomous navigation has become of increasing prominence. Companies including Uber,  Google, and Tesla have worked on and tested their state-of-the-art self-driving cars. This space has vigorously changed over the last decade. One important constituent of the 'self-driving' problem is motion planning, breaking down a desired action into discrete motions that satisfy movement constraints and optimize movement to avoid collisions. There are other opportunities for self-navigating vehicles in non-road settings, including social robots that can navigate crowded spaces. Applicable examples are driving through malls, walking dogs through populated parks, or helping blind pedestrians navigate around others. In such spaces with a non-trivial density of human activity, it's crucial for robots to navigate crowds organically. Humans have multiple unconscious rules when interacting with others' trajectories. They adopt numerous common sense rules and comply with social convention. For example, they plan where to move next by considering their immediate neighbors' personal spaces as well as understanding who has the right-of-way. So, an autonomous robots should also adopt a similar model of movement that fluidly circumvents humans by predicting their trajectories and roughly complying with the same common sense rules as those they share space with. This problem is a good application of recurrent neural nets. The solution will be that of a sequence generation problem.Related Works