CAREER: Capability Inference and Planning for Teams of Learning-Enabled Agents

About This Grant

Automated agents, such as self-driving vehicles, drones, and robotic assistants, have the potential to transform transportation, delivery services, health care, and various other industries by offering more efficient and safer services. These applications are powered by advances in hardware, artificial intelligence (AI), and machine learning that are leading to a proliferation of new agent capabilities to sense their surroundings more accurately, see and understand the world better, navigate and maneuver with dexterity, and engage in complex reasoning. As AI continues to make rapid progress, these capabilities will also give rise to an increasing number of designs and functions of automated agents tailored to specific tasks. The expected increase in capabilities has the potential to enable automated agents to carry out highly complex tasks, significantly improving safety, reliability, and efficiency across various industries. However, one of the key challenges is ensuring that agents with different capabilities can work effectively as a team. Toward this objective, the project focuses on how to systematically characterize the capabilities of agents, assign tasks to a large team of agents with different capabilities, and what actions the agents should take so that the team completes the overall mission as safely and as efficiently as possible. For example, knowing how well delivery vehicles navigate a city during the day or at night and on large roads or narrow streets would allow users to anticipate the volume of deliveries the fleet can handle logistically and cost effectively. Formally, characterizing the capabilities of agents would allow effective matching of agents to tasks, efficient overall planning for teams of agents, and detection of problems during their deployment. The project focuses on reasoning about the capabilities of agents with learning-based components and planning for teams of agents. It will enable the integration of automated agents into existing planning and decision-making frameworks that include both traditional and learning-based components. Approaches that rely exclusively on machine learning are prohibitively expensive since they require large amounts of data, time and energy to train, and are unsafe due to their opaque nature that precludes explanations of their behavior. This work addresses planning with learning-enabled agents by providing formal guarantees for the capabilities and performance of agents. Capabilities are determined by on-board hardware and software, while performance is dependent on the time, location and situation of deployment. Reasoning about agent capabilities enables the integration of learning-based and conventional planning components, provides guarantees and interpretability on agent behavior, and increases effective and efficient use of agents. The three main tasks of the project are developing: (1) A formal framework to describe, reason about and learn agents’ capabilities and performance; (2) Planning methods for teams of agents that match agents to complex tasks based on temporal, spatial and semantic contexts; and (3) Mechanisms for the detection and recovery from failures during deployment stemming from errors in capturing capabilities and estimating their performance. Additionally, the educational activities will prepare students for careers in the interdisciplinary field of planning and decision-making for learning-enabled agents. 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.