CBI-backed NSF project advances AI with open knowledge resource

Whether in robotics or language models, today’s AI systems often falter when faced with real-world tasks requiring deep understanding of complex scenarios and flexible decision-making. This shortfall poses barriers to deploying AI in sensitive domains such as healthcare, government, and manufacturing, where mistakes carry serious consequences.

A new project, recently awarded by the National Science Foundation (NSF) and set to begin on October 1, takes direct aim at this challenge. The initiative will develop an open knowledge resource designed to integrate seamlessly with modern neural AI systems. By integrating reliable knowledge modules into large models, this resource will serve as a bridge between raw data and reasoning capability, making AI more trustworthy, interoperable, and aligned with human standards. The work brings together researchers at Wright State University, Bosch, and Knowledge Systems Research LLC. The Carnegie Bosch Institute will support two collaborators at CMU: Jean Oh (Robotics Institute), focusing on robotics for manufacturing, and Pradeep Ravikumar (Machine Learning Department), focusing on root cause analysis of anomalies in production lines.

The novel knowledge resource will benefit engineers designing agentic workflows and generative AI applications by making it as easy to integrate knowledge modules as it is to install a Python library to add reasoning capabilities. Beyond usability, this resource will serve as a new class of training input for machine learning systems. Whereas most existing datasets capture only raw facts, the knowledge resource will emphasize conceptual knowledge: structured understanding that supports reasoning, generalization, and abstraction. This is especially critical for applications requiring world models, such as robotic planning, which must connect sensor-level data to higher-level cognition, and fact-checking systems, which need to detect false information.

The NSF project advocates a neuro-symbolic approach that blends the strengths of statistical learning with structured reasoning. Rather than relying solely on ever-larger models and training sets, the project aims to extend AI with conceptual knowledge—knowledge that is scalable not through data volume alone, but through strategic collection and abstraction drawn from real-world tasks. Bosch, with its extensive footprint in the American and global manufacturing sectors, is uniquely positioned to contribute valuable, context-rich domain expertise to this effort.

Ultimately, this project represents a step toward “Physical AI”—intelligent systems that can operate robustly in the real world by leveraging both data-driven learning and structured knowledge. By combining perception, abstraction, and reasoning, such systems will be better equipped to make trustworthy decisions in critical domains, supporting safer and more effective human-AI collaboration.