Dynamic representations for autonomous driving

Abstract

This paper presents a method for observational learning in autonomous agents. A formalism based on deep learning implementations of variational methods and Bayesian filtering theory is presented. It is explained how the proposed method is capable of modeling the environment to mimic behaviors in an observed interaction by building internal representations and discovering temporal and causal relations. The method is evaluated in a typical surveillance scenario, i.e., perimeter monitoring. It is shown that the vehicle learns how to drive itself by simultaneously observing its surroundings and the actions taken by a human driver for a given task. That is achieved by embedding knowledge regarding perception-action couplings in dynamic representational states used to produce action flows. Thereby, representations link sensory data to control signals. In particular, the representational states associate visual features to stable action concepts such as turning or going straight.