MINES ParisTech CAS - Centre automatique et systèmes


Séance du jeudi 8 juin 2017, Salle L106, 10h30-12h30.

11h30 - 12h30 Aaron D. Ames, California Institute of Technology, USA.

Humans have the ability to locomote with deceptive ease, navigating everything from daily environments to uneven and uncertain terrain with efficiency and robustness. With the goal of achieving these capabilities on robotic systems—ranging from legged robots, to robotic assistive devices to wheeled and aerial vehicles—this talk will present a unified formal framework for realizing dynamic behaviors in a provably correct and safety-critical fashion, along with the application of these ideas experimentally on a wide variety of robotic systems.

Beginning at the level of behavior synthesis, hybrid dynamical system models will be utilized in conjunction with efficient offline optimization problems to generate virtual constraints that encode the desired behavior of robotic systems. This automatically yields control Lyapunov functions (CLFs) that provably achieve this desired behavior. Going beyond explicit feedback control strategies, these CLFs naturally result in an optimization-based control methodology that dynamically accounts for physical constraints while being implementable in real-time. This sets the stage for the unification of control objectives with safety constraints through the use of a new class of control barrier functions (CBFs) formally enforcing these constraints. These concepts will be illustrated through their application to the humanoid robot DURUS, with the result being dynamic and efficient locomotion displaying the hallmarks of natural human walking: heel-toe behavior. The translation of these ideas to robotic assistive devices, and specifically powered prostheses, will be described in the context of custom-built hardware. Finally, the extension of these concepts to safety-critical systems—including automotive applications, multi-agent systems, and swarms of quadrotors—will be discussed. Therefore, this talk will explore a formal approach to achieving dynamic behaviors on robotic systems, together with the validation of these concepts experimentally on hardware platforms.