Adaptive, Learning Control for Unmanned Vehicles

When using autonomous systems in unknown environments, it will not be possible to anticipate all scenarios in which the devices will be involved and pre-program their response. It will be increasingly incumbent on the machines to ascertain the intent of the users and be able to learn from user actions to improve performance. Intelligent Automation Inc. (IAI) and the University of Pennsylvania (Penn) will explore several new algorithms that would enable a machine to learn some of the high-level strategies that humans use to control and monitor robotic devices remotely. By collecting data from human teleoperation, we propose learning a mode controlling mixture model that would approximate the decision boundaries that human operators use to switch strategies. This idea will take the skill and intelligence inherent in human operators and provide the basis for how the controller can adjust during autonomous operation. In order to optimize the parameters of the controllers, we propose using optimization techniques inspired from genetic algorithms (GAs) to find globally optimal parameters within these modes. To further optimize the learned mode switching behavior, we also plan to incorporate online reinforcement learning (RL) techniques to adaptively tune the learned parameters.