CAREER: Shaping Embodied Intelligence Through Language-Guided Introspection

Artificial intelligence (AI) systems are increasingly deployed in real-world physical environments to assist in factories, hospitals, warehouses, and other dynamic settings, where safety, reliability, and effective human-AI collaboration are essential. Yet current AI agents still fail in ways that are difficult to detect, explain, and correct. They often misinterpret instructions, fail to recognize when their own reasoning is flawed, struggle to recover from errors, and persist in unsafe or ineffective behavior. This project addresses these limitations by developing a new generation of AI agents that can monitor their reasoning, draw on past experience to evaluate and explain their decisions, and adapt their internal understanding of the world based on corrective feedback. Such self-regulating AI agents will be more reliable, more transparent, and safer to deploy in complex physical environments. Beyond advancing the science of AI, this project translates research into learning experiences, student mentoring, interactive demonstrations, and community engagement activities that expand AI literacy and strengthen AI workforce readiness. This project develops a unified framework for embodied intelligence in which language serves as a cognitive mechanism for self-monitoring, self-evaluation, and self-regulation. First, the research develops introspective reasoning methods that enable agents to detect, attribute, and mitigate perceptual and reasoning failures during task execution. Second, the research develops reflective explanation mechanisms grounded in episodic memory, enabling agents to retrieve relevant past experiences and their associated reasoning to assess ongoing decisions and proactively seek targeted human feedback. Third, the research develops language-guided self-organizing world models that enable agents to revise their internal beliefs and dynamically adapt decision-making in response to evolving feedback and introspective diagnostics. The project will produce new algorithms, benchmarks, evaluation protocols, and openly shared resources for reliable embodied AI. By advancing our scientific understanding of the fundamental principles underlying intelligent embodied behavior, this work establishes new foundations for how language enables self-regulation and adaptive decision-making in physical environments. 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. NSF Award ID: 2542328 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT,01002930DB NSF RESEARCH & RELATED ACTIVIT,01003031DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Ismini Lourentzou | Institution: University of Illinois at Urbana-Champaign, URBANA, IL | Award Amount: $356,787 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2542328 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2542328.html