Making Sense: Simultaneous Sensor Configuration and Optimal Control for Autonomous Systems

Worcester Polytechnic Institute, Worcester MA

Investigators

Abstract

This project will contribute new scientific knowledge for future autonomous systems in large-scale applications, including emergency response to adverse events such as natural disasters or industrial accidents. As an example, in a post-disaster scenario where many roads are flooded, it is important to determine viable evacuation routes to transport people to safer locations. It is desirable to automate this process as much as possible: merge information from various sources and quickly plan the safe route. A network of unmanned aerial vehicles can visually survey the extent of flooding and identify passable roads. Information about road conditions, downed power lines, and other environmental factors may also be gained from diverse sources such as traffic cameras or social media posts. The current science of autonomy is inadequate because information-gathering and planning are typically treated as two separate problems. By contrast, this project emphasizes simultaneous information-gathering and planning, namely, methods to identify and deploy the most relevant sources of information in the context of a specific planning objective. With this approach, optimal plans may be achieved significantly faster than is possible with current technology. Consequently, the outcomes of this research may enable faster responses to adverse events, which in turn may help reduce loss of life and damage to property and the environment. In new autonomous systems, sensors may be exteroceptive, multimodal, and configurable, e.g., parameters such as location and pan-tilt-zoom may be tuned. This research aims to close the loop between estimation and control by configuring sensors to collect data most relevant to an optimal control objective. This is a departure from the separation principle traditionally used in control design. With such an integrated approach, it is anticipated that the control objective can be achieved with less risk and significantly lower volumes of sensor data, in some cases orders of magnitude fewer measurements, compared to the traditional paradigm. The research combines new machine learning tools with control and estimation techniques to execute an iterative sensing and control algorithm. At each iteration, an optimal sensor configuration is achieved for a given control design by maximizing a context-relevant information gain metric, and an optimal control design is then determined from the data generated by the updated sensor configuration. The fixed point of these iterations is a near-optimal solution to the coupled problem. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.