Composite Adaptive Predictive Safety via Barrier-Constrained Optimization

This project aims to make autonomous machines safe while they learn. Drones that inspect storm damage, mobile robots that navigate warehouses, and automated vehicles must respect strict physical limits even though their internal mathematical models never match the real world exactly. Current control methods force an unwelcome choice: a machine can act cautiously enough to remain provably safe, often failing to complete its task, or it can learn and adapt quickly while risking a safety violation during the learning process. The project will bring transformative change by removing this trade-off, so that a machine can plan ahead, seek out informative experiences, and refine its own model without ever leaving a certified safe operating envelope. This will be achieved by unifying three elements in one mathematical framework: an adaptive safety filter, a forward-looking motion planner, and a learning module that quantifies its own uncertainty. The intellectual merit of the project includes new mathematical conditions under which safety, prediction, and learning can be certified jointly rather than separately. The broader impacts of the project include an open-source software toolbox for researchers and industry, research training for graduate and undergraduate students, and hands-on robotics workshops for high school students on the Gulf Coast. The technical aim of this project is to develop the composite adaptive predictive safety framework, an architecture that couples online parameter adaptation, receding-horizon optimization, and statistical learning within a single safety certificate. The foundation is a composite energy function that merges a control barrier function, which encodes the safety constraint, with a control Lyapunov function and a parameter estimation error term. The first research thrust employs this energy function as the terminal cost of a nonlinear model predictive controller and seeks computable conditions under which the optimization remains feasible at every step while parameter estimates evolve. The second thrust converts the classical persistent excitation assumption of adaptive control into a designed constraint on the Fisher information accumulated along planned trajectories, so the controller deliberately gathers informative data while a high-frequency safety filter enforces constraints. The third thrust extends the framework to uncertainty without parametric structure by embedding the posterior variance of a Gaussian process, a data-driven statistical model, into the energy function, allowing safety margins to contract as data accumulate. Validation proceeds from simulation benchmarks to hardware experiments on a ground robot under friction uncertainty and a miniature quadcopter under spatially varying wind. 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.