SBIR Fast-Track: Scalable Robotic Skill Learning via Generative Embodied Twins and Vision-Language-Action Models for Task-Oriented Autonomy

The broader impact of this Small Business Innovation Research (SBIR) Fast-Track project is safe, scalable, and intelligent robotic assistance for high-frequency physical work in commercial buildings. The commercial potential is substantial because these workflows are ubiquitous and repeatable, enabling scalable multi-site adoption. The innovation also advances scientific and technological understanding by developing new methods to translate recent AI and robotics advances into real-world autonomy: how to construct executable digital representations of real scenarios from operational data, how to train and stress-test robot skills across different sites; and how to maintain reliability and safety as systems evolve. This Small Business Innovation Research (SBIR) Fast-Track project will develop a world-model-enabled data engine and learning system that enables robots to execute long-horizon, high-frequency service workflows reliably across many similar real-world sites. The research objective is to develop long-horizon robot skill while reducing site-specific programming and operational intervention by integrating a site-customizable generative world model with a hierarchical vision-language-action (VLA) architecture augmented by operational memory. The proposed research will (1) collect facility scans and robot interaction logs to build and continuously refine an executable world model that generates physically valid, task-relevant scenario distributions for training and stress testing; and (2) train hierarchical VLA policies that combine learned skill primitives with demonstrations and real rollouts, while incorporating verification, recovery, and human-in-the-loop supervision to maintain safety and consistency over time. 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.