|Lectura Tesi Doctoral Pragna Das|
Difusió de les dades de la Lectura de la Tesi Doctoral de : Pragna Das
Titol: ADAPTIVE MULTI-ROBOT CONTROL THROUGH ON-LINE PARAMETER IDENTIFICATION AT SYSTEM LEVEL.
Director/a: Lluís Ribas Xirgo
Data i Hora de la lectura: 16/05/2018 a les 10:00h
Lloc: seminari Q3/0007 de l’Escola d’Enginyeria
Programa de Doctorat: Enginyeria Electrònica i de Telecomunicació
Departament: Microelectrònica i Sistemes Electrònics
Industrie 4.0 is the fourth industrial revolution and is related to the application of the generic concept of cyber-physical systems . The outcome of many projects aligned with Industrie 4.0 proposal includes multi-robot systems (MRS) for manufacturing which are more adaptable, decentralised, service oriented and have real-time control and modularity . One of the most crucial aspects of an MRS is the organization of the flow of control decisions. The control and planning have to accomplish different tasks like order assignment, mitigating collision and maintaining communication among all the robots .
The states of the batteries, robots and environment change often in autonomous systems like MRS. The time required to complete tasks or performance times vary according to the change in states of batteries, floor and mechanical parts of the robots.
This work focuses on identifying these performance times and model them so as to reflect the states of the batteries and environment. The inclusion of these performance times in planning and control decision can produce more cost-efficient decisions. They are modelled using nonlinear state dependent modelling techniques.
Also, they are estimated using the model developed in this work and the values are used in the decision-making process. This is done to study their efficacy in planning in an MRS used for internal transportation.
A prototype MRS for logistics is prepared for the experiments where the floor is described by a topological map with nodes designating ports or junctions and edges connecting the nodes. The robots carry materials from one port to another.
The paths between ports are composed of different edges. Traversing an edge is considered as a task. The travel time spent by a robot to traverse an edge is the designated performance time which indicate the cost involved to traverse the edges.
Thus, travel times reflect continuous changes in robot and environment and influence to modify decisions which are made without considering them. Travel times for a robot are made available online through estimation. Online estimation demands a suitable formalization and model for these travel times. In this work, a state-dependent bilinear model from time-series modelling technique is used to model travel times in each mobile robot in order to estimate them online. The efficacy of these travel times is studied by complementing them into a planning algorithm. Routes are computed to find the shortest path from one port to another for each MR. The travel times i are used as weights of edges into route planning instead of weights derived from heuristics or other cost factors. In route planning for a single robot, the travel time for an edge may be required to be estimated for multiple times. The estimated travel times between any two nodes provide the current and close-to-real cost of traversing at different instances. These estimated values form a profile of travel costs for edges through the duration of operation of the robot. These estimated travel times are different than heuristics costs as they depict the real states which are impossible to know from heuristics. This facilitates path planning algorithms to choose the edges with least real travel times or costs to form the path. The experiments show that path obtained through online estimated travel times are of 15% less total cost compared to that obtained by heuristics costs.
Nevertheless, a good estimation is dependent on historical data which are close in time. But, there are situations when all the travel times for one or more edge(s) are not available for the entire duration of operation of the MRS to an individual robot.
The proclivity of this occurrence lies in the fact that the edge may not have been travelled even once by the robot, or travel time for that edge have not been recorded in recent past.
Then, it is imperative for that robot to gather the necessary travel times from others in the system as a reference observation. But, these observations are from other robots in different battery condition than itself. Still, the bi-linear model for travel time for the robot itself using other robots’ observation and its own change or exploration in the travel times till the current instance. The crux of this process is to predict current travel times in the robot using others’ travel time for the same edge.
The mechanism of information sharing between one robot to others in the system has been devised in a form of a common ontology-based knowledge. This ontology structure is identical in each robot which contains the travel times of edges with contexts attached to each data about the instances of estimation, the nodes that particular edge connects and other pieces of information. This ontology helps to fetch and share information forming a collective knowledge base facilitating a comprehensive control and planning for the system.
This greatly helps the MR to estimate travel times more accurately and precisely.
Also, accurate estimation affects route planning to be more precise with reduced cost.
The total cost of paths generated through the travel times estimated through sharing is 40% less on average than that of paths generated through travel times without sharing.
In this work, only a single task is considered whereas in a real industry a robot needs to do a variety of tasks. This work paves the way to consider all the varieties of tasks in an automated system and identify different types of cost coefficients, other than travel times. In that case, estimating and sharing information would be in a bigger domain with more complexity which demands artificial intelligence to be used along with re-enforcement learning. The problem domain can be further enhanced with different kinds of robots in a system like unmanned aerial vehicles, other ground vehicles, and human agents.
|Anar a Notícies|