DeepTaxi: Teaching an Autonomous Car with Social Scaffolding

In this paper, we present DeepTaxi, an extension to an existing autonomous RC car platform that allows for the dynamic learning of an environment. DeepTaxi employs a social scaffolding approach where a human user supervises and initially provides feedback to the car so that it can learn the names and order of various objects located around a track. Once it sufficiently learns about the environment, DeepTaxi can then autonomously navigate to any desired location without the need for human assistance. We test DeepTaxi with human participants on a custom made track with a variety of objects/orders. We find that it can successfully learn about and navigate the track with the participants expressing appreciation for the timeliness of the car's communication.