MODEL PREDICTIVE CONTROL AND IMITATION LEARNING ALGORITHMS FOR ROBOT MOTION PLANNING IN PHYSICAL HUMAN-ROBOT  INTERACTION

Aigerim Nurbayeva

MODEL PREDICTIVE CONTROL AND IMITATION LEARNING ALGORITHMS FOR ROBOT MOTION PLANNING IN PHYSICAL HUMAN-ROBOT INTERACTION

dc.contributor.author	Aigerim Nurbayeva
dc.date.accessioned	2024-10-15T10:16:09Z
dc.date.available	2024-10-15T10:16:09Z
dc.date.issued	2024-08-07
dc.description.abstract	This PhD thesis focuses on the design and testing of safe robot motion planning algorithms for human-robot workspace sharing. These algorithms are based on the use of nonlinear model predictive control (NMPC), a model-based method for motion planning relying on numerical optimization. The contribution of the thesis can be split into two main areas. The first area consists of the approximation of NMPC laws using deep neural networks (DNNs), often referred to as “imitation learning”. This is motivated by the fact that the execution of NMPC laws might require a considerable amount of time, which restricts the performance of the closed-loop system. Calculating the output of a DNN for a given input is instead a much faster process. Therefore, replacing the optimization solver of NMPC with a DNN can reduce computation times, thus improving performance. It is crucial, though, to suitably train the DNN to imitate the NMPC law in order to improve performance and at the same time guarantee safety. The final result obtained in this area consists of using the so-called dataset-aggregation approach for DNN training, together with properly designed safety filters, which ensure that the safety constraints imposed in the NMPC problem also hold for the robot motion generated by the DNN. The second area consists of the extension of a previously defined NMPC law in terms of stabilizing terminal constraints. The most common approach for guaranteeing closed-loop stability in an NMPC problem is the imposition of terminal constraints, i.e., the prediction of the system motion is required to satisfy certain conditions at the end of the prediction horizon. Specifically, in a previous approach, the “point terminal constraint” was used, in which the prediction of the robot motion had to exactly reach the desired goal configuration at the end of the prediction horizon. In this thesis, this condition is relaxed by imposing that a given set, rather than a given point, is reached in the state space for the predicted robot motion. The imposition of this new condition allows for an enlargement of the domain of attraction, i.e., the NMPC law can find a solution for reaching the goal configuration from a wider set of initial configurations. All the proposed motion planning strategies were tested experimentally on a UR5 collaborative manipulator.
dc.identifier.citation	Nurbayeva, A. (2024). Model Predictive Control and Imitation Learning Algorithms for Robot Motion Planning in Physical Human-Robot Interaction. Nazarbayev University School of Engineering and Digital Sciences
dc.identifier.uri	https://nur.nu.edu.kz/handle/123456789/8280
dc.language.iso	en
dc.publisher	Nazarbayev University School of Engineering and Digital Sciences
dc.rights	Attribution-NonCommercial-ShareAlike 3.0 United States	en
dc.rights.uri	http://creativecommons.org/licenses/by-nc-sa/3.0/us/
dc.subject	Type of access: Open access
dc.subject	nonlinear model predictive control
dc.subject	deep imitation learning
dc.subject	physical human-robot interface
dc.subject	motion planning
dc.subject	deep neural networks
dc.subject	stabilizing terminal constraints
dc.subject	UR5
dc.title	MODEL PREDICTIVE CONTROL AND IMITATION LEARNING ALGORITHMS FOR ROBOT MOTION PLANNING IN PHYSICAL HUMAN-ROBOT INTERACTION
dc.type	PhD thesis