Control Engineering Practice 101 (2020) 104488 Contents lists available at ScienceDirect Control Engineering Practice journal homepage: www.elsevier.com/locate/conengprac Robot control parameters auto-tuning in trajectory tracking applications Loris Roveda ∗ , Marco Forgione, Dario Piga Istituto Dalle Molle di studi sull’Intelligenza Artificiale (IDSIA), Scuola Universitaria Professionale della Svizzera Italiana (SUPSI), Università della Svizzera italiana (USI), via Cantonale 2C- 6928, Manno, Switzerland ARTICLE INFO Keywords: Controller auto-tuning Bayesian optimization Trajectory tracking Dynamics compensation Industrial robots ABSTRACT Autonomy is increasingly demanded to industrial manipulators. Robots have to be capable to regulate their behavior to different operational conditions, adapting to the specific task to be executed without requiring high time/resource-consuming human intervention. Achieving an automated tuning of the control parameters of a manipulator is still a challenging task, which involves modeling/identification of the robot dynamics. This usually results in an onerous procedure, both in terms of experimental and data-processing time. This paper addresses the problem of automated tuning of the manipulator controller for trajectory tracking, optimizing control parameters based on the specific trajectory to be executed. A Bayesian optimization algorithm is proposed to tune both the low-level controller parameters (i.e., the equivalent link-masses of the feedback linearizator and the feedforward controller) and the high-level controller parameters (i.e., the joint PID gains). The algorithm adapts the control parameters through a data-driven procedure, optimizing a user-defined trajectory-tracking cost. Safety constraints ensuring, e.g., closed-loop stability and bounds on the maximum joint position error are also included. The performance of proposed approach is demonstrated on a torque- controlled 7-degree-of-freedom FRANKA Emika robot manipulator. The 25 robot control parameters (i.e.,4 link-mass parameters and 21 PID gains) are tuned in less than 130 iterations, and comparable results with respect to the FRANKA Emika embedded position controller are achieved. In addition, the generalization capabilities of the proposed approach are shown exploiting the proper reference trajectory for the tuning of the control parameters. 1. Introduction 1.1. Context Nowadays, robots are required to adapt to (partially) unknown situations, being able to optimize their behaviors through continu- ous interactions with the environment. In such a way, robots can achieve a high level of autonomy which allows them to face unforeseen situations (Makridakis, 2017). Such capabilities are in particularly needed by industrial manipulators, which are expected to execute highly dynamic tasks (Thoben, Wiesner, & Wuest, 2017; Wong, Yang, Yan, & Gu, 2017). This requires reconfigurability, adaptability and flexibility. Indeed, the manipulator has to autonomously adapt to new tasks and working conditions, avoiding as much as possible the human intervention. In fact, human intervention can be time- and resources- consuming and not always feasible (e.g., for safety reasons) (Bruzzone, Massei, Di Matteo, & Kutej, 2018; Pichler et al., 2017; Starke, Hahn, Pedroza Yanez, & Ugalde Leal, 2016). In order to achieve this autonomy, the manipulator has to be capable to self-adapt for the task at hand. Machine learning techniques are extremely effective to tackle such a problem, and many approaches ∗ Corresponding author. E-mail addresses: loris.roveda@idsia.ch (L. Roveda), marco.forgione@idsia.ch (M. Forgione), dario@idsia.ch (D. Piga). have been investigated in recent years to improve the level of autonomy of manipulators through auto-tuning methodologies. 1.2. State-of-the-art machine learning techniques for control tuning Controller design and tuning is one of the most investigated topics in robotics. Standard model-based methodologies require identification of the robot dynamics, with data gathered from time-consuming ad- hoc experiments (Jin & Gans, 2015; Swevers, Verdonck, & De Schutter, 2007). Furthermore, when estimating a model of the manipulator, it is hard to determine a priori the model accuracy required to meet a given closed-loop performance specification. Direct methods, such as Iterative feedback tuning (IFT) (Hjalmarsson, Gevers, Gunnarsson, & Lequin, 1998) and Virtual reference feedback tuning (VRFT) (Formentin, Piga, Tóth, & Savaresi, 2016; Lecchini, Campi, & Savaresi, 2002), could be used to design a controller based on open-loop data, without identifying a model of the system. However, these methodologies require to a- priori specify a model of the desired closed-loop behavior, which might not be achieved by the chosen controller structure. Although some tech- niques have been recently proposed in Novara, Formentin, Savaresi, https://doi.org/10.1016/j.conengprac.2020.104488 Received 19 September 2019; Received in revised form 20 April 2020; Accepted 28 May 2020 Available online 8 June 2020 0967-0661/© 2020 Elsevier Ltd. All rights reserved.