Robotic systems rarely admit precise first-principles models. We develop frameworks that learn control-relevant structure directly from data — principally through operator-theoretic approaches that transform nonlinear dynamics into globally linear representations amenable to optimal and predictive control. A key innovation is closing the loop on this learning: the learned model adapts online to parametric variations, disturbances, and sensor noise, while the closed-loop dynamics remains provably linear.

Deploying robots in the wild demands navigation strategies that are general across platforms and robust to real operating conditions. We pursue two routes: (i) scale-aware deep reinforcement learning that trains a single policy over an entire class of UAS platforms, and (ii) bioinspired optic-flow guidance that recovers the signals needed for safe landing and obstacle avoidance directly from raw visual streams — no mapping, no state estimation.

Real manipulation tasks demand controllers that simultaneously satisfy joint limits, input saturation, safety boundaries, and timing requirements — without model knowledge. We design model-free prescribed-performance controllers for Euler-Lagrange systems, where the user specifies transient and steady-state behavior in advance and the controller guarantees it. Fault-tolerant extensions handle joint motor failures gracefully.

Controllers can only act on what they can estimate. We design nonlinear observers that recover critical signals — distances, velocities, structural scene parameters — from noisy, informationally sparse sensor streams. Concurrent learning relaxes classical persistent-excitation requirements, enabling convergence even when the scene provides intermittent information, which is the norm in outdoor visual navigation.