Safety of Embodied AI and Autonomous Systems: Control, Safe Learning, Verification, and Runtime Assurance

When AI acts through a physical system, an error is no longer only an incorrect output. It becomes a force, a trajectory, a collision risk, or a decision whose consequences may arrive before a human can intervene.

Safety of Embodied AI and Autonomous Systems develops the mathematical and engineering foundations needed to reason about those consequences. The book moves from hybrid safety contracts, hazard sets, physical loss, reachability, and viability to Lyapunov methods, control barrier functions, constrained action selection, stochastic reach-avoid analysis, chance constraints, and risk measures. It then extends the framework to learning-enabled systems through belief-space safety, safe exploration under uncertainty, and learned components operating off distribution.

A unifying theme is the intervention margin: the time or state-space room available before a hazard becomes unavoidable, measured against sensing, computation, communication, actuation, and human-response delays. This perspective carries through runtime assurance and backup controllers, shared autonomy, fault-tolerant operation, degraded modes, and the final verification and safety-case framework.

Designed for graduate students, researchers, and engineers in AI safety, robotics, control, and autonomous systems, the text combines formal results with engineering interpretation, computational labs, structured evidence, and safety-case templates—showing how control, learning, verification, and runtime assurance must fit together when AI decisions have physical consequences.