Fast and Reliable Online Retraining and Adaptation for Robot Planning Despite Missing World Knowledge

George Mason University, Fairfax VA

Investigators

Abstract

The next generation of service robots will be required to act in unfamiliar and ever-changing environments. At the request of human operators, such robots will be expected to reliably complete complex objectives despite missing or out-of-date information about their surroundings: locating key places, delivering supplies, and finding personnel, even when they are uncertain where to look. While machine learning has proven an important component of good behavior in this domain, learning-driven strategies can be brittle to change, resulting in poor performance in new or unfamiliar environments with little recourse to improve without significant downtime and supervision. Poor performance begets mistrust, limiting the adoption of service robots and thus their potential to provide autonomy or assistance to human operators in settings ranging from homes to hospitals. This project aims to overcome these limitations through development of an approach for service robot decision-making designed to allow state-of-the-art performance in challenging, unfamiliar environments and facilitate fast and reliable improvement during deployment. Our contributions will allow for more performant, reliable, and trustworthy robots capable of good behavior in unstructured environments. One key aspect of our work will facilitate non-expert users to quickly correct robot behavior, a capability unique to our proposed approach in this domain that will help democratize robot training, a step towards more trustworthy and ethical robots. Moreover, our advancements will help to lower the barrier to entry for student engagement with robotics and machine learning and our research program is integrated with educational initiatives that engage both undergraduates and students from D.C. area high schools. Our project will develop a principled approach for improving robot behavior during deployment for long-horizon planning in partially-mapped environments, emphasizing reliability, data efficiency, and performance. We will demonstrate that the coupling of data-driven (learning-informed) and classical (STRIPS-style) planning afforded by our abstraction will be a key enabler of this advance; learning will augment model-based planning, allowing completeness and introspection despite missing knowledge. Our robot will rely on two complementary sources of information: (i) online experience, in which the robot uses data it collects during deployment to self-audit and to retrain and adapt its learned behaviors, and (ii) expert guidance, in which an environment expert (e.g., a robot or human auditor) intervenes to prompt a change in long-horizon behavior. Our project will build upon recent progress in planning under uncertainty, trustworthy AI, robust planning, and domain adaptation, and therefore has the potential to advance the state-of-the-art in multiple areas at once. We will demonstrate both simulated and real-world experiments in which a mobile manipulator robot must navigate large-scale, unfamiliar home- and hospital-like buildings to complete complex multi-stage tasks involving locating key places, interacting with the environment, and retrieving objects and persons. We intend to both theoretically justify and demonstrate empirically the utility of our proposed approach to quickly and reliably improve deployment-time performance for a variety of service robot tasks. 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.