Intrinsic Robust Planning for Uncertain Robots

Jeudi 5 décembre 2024 de 15h30 à 16h30, Amphi Costes.

Paolo Robuffo Giordano, Directeur de recherche CNRS à l'IRISA

Abstract:

Uncertainties in a robot/environment can have many sources and can be modeled at different levels of detail. For instance, a typical choice in many robotics applications is to consider a generic additive “process noise” (or “actuation noise”), often taken as Gaussian random variables. This additive term can capture the presence of some uncertainty in the robot’s motion/actuation, but its genericity does not allow, in general, to obtain precise predictions of how the actual uncertainties will affect the robot’s motions or actions.

However, in many cases of interest a model of the robot/environment can be considered available with the main source of uncertainty lumped in the inaccurate knowledge of some model parameters. In these cases, the uncertainty is not generic but it has a very specific structure which, if exploited, can lead to better predictions of its effects on the robot motion.

In light of these considerations, in this talk we will review the recent notion of "closed-loop state sensitivity matrix": this quantity locally captures how deviations in the model parameters (w.r.t. their nominal values) affect the evolution of the robot/environment states in closed-loop, i.e., by also taking into account the strengths/weaknesses of the particular control action chosen for executing the task. A norm of the state sensitivity can, for instance, be minimized for generating reference trajectories that result by construction minimally sensitive to parametric uncertainties, thus increasing the intrinsic robustness of their tracking in closed-loop. Furthermore, it is also possible to leverage the sensitivity matrix to obtain time-varying bounds (tubes) on the state and/or input evolution assuming a (known) range of variation for the parameters. This makes it possible to plan robust trajectories that, at least locally, ensure the feasibility of the resulting motion (against state/input constraints) also in the presence of parametric uncertainties.

The talk will revisit the formalization of the closed-loop sensitivity and derived quantities, and will illustrate how these notions can be applied to offline and online robust trajectory planning for several robotic case studies of interest (drones, manipulators).