LIDS Seminar: Aaron D. Ames

"Foundations for Safe Autonomy: Why Learning Needs Control" 

With the rise of humanoids and the rapid deployment of learning across autonomy stacks, the central question is: how can we trust robots to operate safely around us? Despite impressive performance gains, learning at scale introduces fragility—making safety the key blocker to reliable deployment. This talk outlines the foundations for safe autonomy, coupling learning with the formal guarantees from control theory. The cornerstone of this approach is Control Barrier Functions (CBFs), which encode safety as forward set invariance.  This leads to safety filters: real‑time wrappers that take desired commands—even from black‑box, learning‑enabled components—and minimally modify them (only when needed) to keep the system safe.  These filters naturally sit within layered autonomy stacks, yielding a coherent architecture for trustworthy robots. Finally, I will close the loop with learning: utilizing Lyapunov‑ and barrier‑based reward shaping and shielding to enforce stability and safety during training, not just at deployment.  This framework for safe autonomy is grounded in—and will be illustrated by—extensive experimental results across diverse robotic platforms: ground vehicles, drones, aircraft, legged and humanoid robots.