Federal Grant Award Summary

Georgia Tech Research Corp has received a $600,000 Project Grant from the National Science Foundation's Division of Information and Intelligent Systems under the Computer and Information Science and Engineering program (CFDA 47.070), effective September 1, 2026 through August 31, 2030. This award funds research to develop formal reasoning frameworks and provably correct algorithms for reinforcement learning (RL) systems operating under temporal logic specifications. The project addresses critical safety gaps in RL deployments by creating the first asymptotically convergent direct learning algorithms that operate on temporal logic objectives without translating them into numerical rewards, thereby eliminating reward hacking vulnerabilities that could otherwise lead to dangerous failures in safety-critical applications.

The research will deliver three primary products: (1) theoretical foundations and multi-stage stochastic approximation techniques for direct specification-guided RL; (2) a compositional framework enabling long-horizon task learning with formal correctness guarantees; and (3) an open-source certification platform that generates externally verifiable proof artifacts for arbitrary RL policies using black-box access. These tools are intended to enable safe deployment of RL systems in regulated industries including autonomous vehicles and medical robotics, establish community standards for specification-guided RL research, and train the next generation of researchers at the intersection of formal methods and machine learning. The work will be performed at Georgia Tech's Atlanta campus.