Improving Navigation with Learning from Demonstration


General purpose navigation systems for indoor wheeled robots are often capable of running with little to no oversight by humans, even in busy or cluttered environments such as hotels or hospitals. However, these systems are still not 100% successful, and so when the robot is confused or stuck, often a human operator is required to step in remotely and fix the problem. We see this as a learning from demonstration problem, where the goal is to build off the existing autonomy and achieve that last small percent needed for total independence.