Federal agency name: defense advanced research projects agency (DARPA), defense sciences office (DSO) funding opportunity title: young faculty award (YFA) funding opportunity number:…

The Department of the Interior Departmental Offices Interior Business Center awarded a cooperative agreement totaling $499,805 to the Rector & Visitors of the University of Virginia on September 30, 2025, under the Research and Technology Development program (CFDA 12.910), which supports basic, applied, and advanced research with potential military or dual-use application at educational institutions, nonprofit organizations, and commercial firms.

The University of Virginia will develop an active epistemic planning framework paired with large language models augmented with embodied cognition capabilities for deployment in human-robot applications operating in real-world dynamic settings under uncertainty. The research addresses limitations in how large language models translate reasoning about belief states into strategic decisions and actions, specifically their conservative responses despite access to uncertainty-reduction tools and their lack of the processing constraints that guide human embodied decision-making.

The award runs through September 29, 2027, with performance occurring in Virginia. This cooperative agreement was funded under the Young Faculty Award program solicitation (DARPARA2501), which sought innovative research proposals across DARPA's Defense Sciences Office, Information Innovation Office, Microsystems Technology Office, and Tactical Technology Office, targeting revolutionary advances in science, devices, or systems rather than evolutionary improvements to existing practice.

Federal agency name: defense advanced research projects agency (DARPA), defense sciences office (DSO) funding opportunity title: young faculty award (YFA) funding opportunity number: DARPARA2501CATALOG of federal domestic assistance (cfda) number(s): 12.910 research and technology developmentprogram description: DARPA is soliciting innovative research proposals in the areas of interest to the following DARPA technical offices: defense sciences office (DSO), information innovation office (I2O), microsystems technology office (MTO), and the tactical technology office (TTO). Further detail regarding the specific technical areas of interest can be found below under open office topics. Proposed research should investigate innovative approaches that enable revolutionary advances in science, devices, or systems. Specifically excluded is research that primarily results in evolutionary improvements to the existing state of practice. The goal of this project is to enable embodied cognition capabilities in large language models(llms) deployed in human-robot applications in real world dynamic settings under uncertainties.llms demonstrate sophisticated and accurate performance in reasoning tasks about belief statesand compute mentalistic-like inferences. However, they struggle to translate this reasoning intostrategic decisions and actions and hence perform differently from humans as they do not computethese inferences spontaneously to reduce uncertainty. Resolving such uncertainty is a priority forbrains adapted to deal with embodied decisions. Llms respond conservatively despite havingaccess to tools to reduce uncertainty, because they do not operate within an environment and arethus not subject to the processing constraints that humans face to resolve competition betweenaction choices, so may have limited advantages in narrowing the future prediction space.technical approach. We propose an active epistemic planning (AEP) framework paired withllms augmented with embodied cognition to enable a robot to empathize with a human, providea reasoning structure to anticipate and predict future challenges, and reason and commit spontaneouslyabout potential actions. Specifically, the proposed AEP framework will use second andhigher order theory of mind reasoning through llms to estimate the humans state and their beliefover the robots state and decide what action to perform considering sensed physical and semanticconstraints in the environment and their uncertainties. Because multiple policies can be formed atrun-time, the challenge here is for the robot to choose the best one. To solve such estimation anddecision-making problem, we propose an active inference approach to minimize surprises in thesense that the autonomous system will pick control planning actions that minimize the error in thepredicted observations, hence helping the human-robot team achieve their final objective. Finally, we propose novel explicit communication strategies, risk-aware policies, and conformal predictionmethods to quantify uncertainties and provide safety and robustness guarantees.