CPS: Medium: Information based Control of Cyber-Physical Systems operating in uncertain environments

Northwestern University, Evanston IL

Investigators

Abstract

Most cyber-physical systems have operated in comparatively benign environments, often engineered to meet the needs of the system. For instance, in many settings robots operate in closed-off environments in manufacturing. However, as these systems are deployed in increasingly isolated environments (such as search and rescue efforts, automation in surgical devices, collaborative manufacturing) these robots will need to operate while reasoning about substantial uncertainty. Moreover, actions taken by the system impact that uncertainty. The proposed work will develop algorithms that enable a cyber-physical system to reason about what actions it must take to manage its uncertainty. A simple example of this is understanding that one must look behind and under things when searching for an object. The goal of this project is to automate the process of managing that uncertainty, however it arises. For instance, interacting with humans (such as physically interacting with a person while assisting her motion, or tracking a person during a search effort) involves uncertainty about what the person is going to do. Enabling a robotic system to actively test a person's intent, and then act according to her subsequent behavior, is key to the robot's ability to be an effective partner. This essential point, that action on the part of a cyber-physical system can be used to explicitly manage its understanding of the world, is the core purpose of the proposed work. The proposed work will leverage recent results in information-based real-time nonlinear control. Specifically, ergodic control (controlling the ergodicity of a trajectory relative to some reference distribution) enables one to specify an objective using a density function to describe the desired spatial characteristics for a trajectory. In the context of Hidden Markov Models (HMMs) and Partially Observable Markov Decision Processes (POMDPs), this allows a system to actively probe multiple states, simultaneously considering process uncertainty, measurement uncertainty, and uncertainty associated with the Markov process itself, even when the HMM is changing over time because of new states or processes being introduced into the ambient environment. The apparent reason for the effectiveness of ergodic control in the context of reactive planning under uncertainty is that it always computes plans in the continuum, avoiding the combinatoric complexity associated with POMDPs. Moreover, the needs of the dynamic system and the cyber system are inter-related during physical motion, so the work will additionally develop methods that maintain stability with respect to both state and information. This inter-dependence creates a trade-off between the physical ability to act and the computational load on the cyber system, enabling a cyber-physical system to reduce the load on its cyber system through physical action. The research will develop algorithms capable of real-time information-based control in response to constantly changing data. One of the motivating examples will be aerial vehicles tracking unknown numbers of targets on the ground in a highly occluded environment such as a forest or an urban setting. Additionally, a robotic arm, already used for assistive device research, will be used to implement and assess the proposed methods. 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.