GOALI: Conformalized, Risk-Adaptable Coordinated Control for Autonomous In-hospital DElivery Robots (AIDERs)

Like many public places, hospitals are busy and dynamic. Doctors and nurses spend valuable time moving medications, medical equipment, meals, and supplies from one location to another. While some delivery robots already exist, they are limited in their ability to work safely and efficiently in crowded hallways where people, wheelchairs, stretchers, and unexpected events constantly change the environment. This Grant opportunity for Academic Liaison with Industry (GOALI) project will develop new analytical methods that allow teams of robots to work together, make intelligent decisions, and adapt to changing conditions as they transport supplies throughout a hospital or any other busy and unpredictable work environment. By helping with routine delivery tasks, these robots could free healthcare workers to spend more time caring for patients while improving the efficiency and reliability of hospital operations. The research supported by this grant will establish the foundations on how multi-agent systems (such as multi-robot systems) assess and adapt risk under nominal and emergency conditions, so that they can operate reliably in homes, museums, warehouses and stores, increasing productivity while reducing labor. The work applies more broadly to autonomous systems and robots in transportation, urban air mobility and defense, establishing the foundations of resilient autonomy. The research will also lead to curriculum advancements in multi-agent systems, robotics and control, benefiting engineering education. Multi-agent systems should be able to achieve the following capabilities in real time: (i) Reason about the uncertainty of the predictions of their own states and of their environment, which is captured as AI/ML-modeled disturbances. (ii) Assess the feasibility of their mission, as well as the risk induced by their possible decisions. (iii) If needed, adapt their risk to optimize for mission objectives while meeting minimal mission requirements. These challenges require novel optimization-based planning and control methods. This project will build a framework on assessing constraint feasibility as a novel notion of risk, informed by confidence in the estimates of the uncertain environment and, if needed, adapting this risk online, so that constraints can be relaxed as needed to maintain feasibility under guaranteed minimal system requirements. This framework will be used for seamless operation in crowded, dynamic and confined environments. The project provides insights on how to handle constraint incompatibility as a risk that the system is willing to adjust on-the-fly to meet minimal requirements, informed by the confidence level of the predictions of uncertain multi-agent dynamical systems. Specifically, the project will develop conformal prediction methods for quantifying uncertainty in dynamical, multi-agent systems; assess the risk of the agents’ decisions by accounting for the environmental uncertainty and the compatibility of the mission constraints using tools from constrained optimization and duality theory; and in the case of incompatible constraints, adapt the risk so that it can increase while not violating minimal system requirements, leading to principled relaxation of constraints. 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.