CAREER: Trustworthy Learning-Enabled Autonomy: Safe, Robust, and Scalable Data-Driven Decision-Making

This NSF CAREER project aims to develop foundations that allow autonomous systems to learn from data while reliably respecting safety constraints. AI-enabled vehicles, robots, and infrastructure controllers are moving into public spaces and critical services, where rare mistakes can cause injuries or cascading disruptions. Many modern learning components provide limited guarantees: they may violate constraints, fail when conditions differ from training data, or scale poorly when many decision makers must coordinate. The project will bring transformative change by making safety and reliability a built-in property of learning-enabled autonomy—from individual neural networks, to learning-based controllers, to large multi-agent systems—thereby promoting the progress of science and advancing the national health, prosperity, welfare, and security. This will be achieved by combining hard-constrained learning architectures, run-time uncertainty monitoring, and scalable decentralized decision-making tools. The intellectual merit of the project includes new mathematical theory and efficient algorithms for constraint-satisfying learning, uncertainty-aware and uncertainty-averse decision making, and safe coordination in multi-agent systems. The broader impacts of the project include safer autonomous transportation and robotics, more reliable and energy-efficient operation of engineered systems, open-source tools and datasets, and integrated education and outreach that strengthen K–12 through graduate training and engage the public through interactive demonstrations. Technically, the research comprises three coupled thrusts. Thrust 1 will create Hard-Constrained Neural Networks (HardNet), which add a differentiable projection layer so input-dependent constraints are satisfied by construction during training and deployment. HardNet will be embedded in reinforcement learning (RL—learning by trial and error) to produce certifiably safe policies, and in boundary control of partial differential equations (PDEs—models of distributed physical processes) by enforcing Lyapunov stability conditions as hard constraints. Thrust 2 will develop model-agnostic run-time uncertainty metrics for pre-trained perception and representation models using neighborhood-consistency tests and scalable curvature “sketches” (efficient sensitivity summaries), enabling detection of out-of-distribution inputs without costly retraining. These signals will guide adaptive “uncertainty-aware” and proactive “uncertainty-averse” policies, and robust adaptive safe RL that adjusts online rather than relying on pessimistic worst-case design. Thrust 3 will scale these guarantees to many agents via decentralized safe multi-agent RL that maintains controlled-invariant safe regions using Hamilton–Jacobi reachability (computing safety envelopes), federated RL that personalizes collaboration under heterogeneous dynamics, and online incentive mechanisms that satisfy global constraints under incomplete information. The methods will be validated in robotics, transportation, and building-energy testbeds and disseminated through open educational materials. 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: 2544396 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Navid Azizan Ruhi | Institution: Massachusetts Institute of Technology, CAMBRIDGE, MA | Award Amount: $659,678 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2544396 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2544396.html