Safety Embedded Autonomy: From Barrier States to Intelligent and Differentiable Control

Overview

Autonomous systems must satisfy hard safety constraints while pursuing performance objectives, adapting to changing dynamics, and increasingly learning from interaction. This raises a fundamental question: how should safety enter the control loop?

This talk presents a safety embedded perspective in which safety is incorporated into the dynamics used for control design, rather than treated only as an external constraint or corrective filter. This perspective led to Barrier States (BaS), which transform safety constraints into dynamical states whose boundedness characterizes safety. Crucially, this reformulation turns safety-constrained control into control of an augmented dynamical system, enabling established feedback and optimal control methods to extend naturally to safety-critical design — from classical control to differential dynamic programming for robotic trajectory optimization.

This viewpoint naturally leads to a second challenge: what if the dynamics themselves are uncertain or change during operation? A further question follows: can the system itself learn to improve through physical interaction while respecting its control and safety requirements? These questions motivate safety embedded adaptive control and rapidly adaptive safety filters that infer changing dynamics online, together with differentiable control methods that use task-level gradients to adapt performance, robustness, and safety directly from experience.

Together, these directions point toward safety embedded autonomy in which safety, adaptation, and learning are designed together to enable autonomous physical systems that can operate reliably, adapt to changing environments, and improve through interaction.