According to Chiddarwar and Babu (2017), there are many issues related with safe and fast task operation when multiple robots share a common workspace of agent-based systems, which are widely used in decision making and controlling systems, besides, each task is important for coordinated motion planning of multiple robots in the agent-based framework. However, the need of this flexible and systematic framework is facing the challenge from modelling work units to performing trajectory generation regard to the growing number, enhanced emission’s complexity and constrain number of industrial robots. As a result, designing conventional motion planning system for those multiple robots to devise strategies have to concern the geometry for robot and environment modelling, schemes for collision avoidance and strategies for coordination.

From pervious experience, automatic system is limited by the vision and software that installed by programmers, however, the superior machine learning algorithms are alters from the traditional way. Computational methods are required by machine learning algorithms to affective learn directly from data and without relying on program or human operation (Automation World 2018). Due to the variations of system and task, high performance of robots can be randomly achieved. This is to say, if letting every robot perform consisting studying at customer’s site, they may not be able to collect the necessary input or output data during the learning process because of the limited measurements access in sensing system (Zhu & Wang 2020).