Human workers involved in manufacturing use-cases know which tasks they must execute and in which order, such that they can achieve their goals, e.g. assembly a group. The description of the tasks and their ordering can be presented, for example, in a construction manual. This manual contains information about the steps that must be carried out, the required tools, and possible the time duration needed for each assembly step.
It is possible that more products must be assembled at the same time or that more human workers are involved in a complex assembly process. In these cases, the distribution of work among the involved persons is discussed at the beginning of the day, for example during internal meetings, or follows dynamically, during the processes’ execution. The same approach for the work re-distribution among the involved parties is applied when delays or other errors occur during the assembly.
Human workers master the selection of tasks, their execution, and their re-allocation is case of faults. The workers are able to take all required measures to reach their goals by the end of the day. However, when robotic systems are also involved in these manufacturing processes, the tasks definition, allocation, and execution become much more difficult.
The involved robot is not able to decide by itself which action it can overtake or how it should automatically adapt the tasks execution depending on the actual status of the environment. The robot cannot make decisions about the best way to collaborate and cooperate with the human workers. To tackle this challenge, RWTH deploys Automated Task Planning approaches, also known as AI Planning approaches. AI Planning approaches enable the automatic generation of tasks for the involved actors, considering the actual state of the world, and ensuring that set goals can be reached.
AI Planning approach for the automatic generation of tasks
The AI Planning approach works as follows. For a considered use-case scenario, a planning problem is modelled. Each model contains the abstract description of the actions that the actors can carry out and an abstract representation of the environment. The latest is composed of instances of the agents and of other objects or poses from the environment that are relevant for the planning problem. Such objects are, for example, the elements of the product that must be assembled, while relevant poses are those locations where the human worker or the robot can move to in order to carry out a specific task.
Besides all this information regarding the actions, the agents, and the environment, to complete a planning problem an initial state and a goal state must be defined. The initial state usually describes the actual state of the environment, while the goal state sets what must be achieved. An example of a goal state is that all products due in a specific day must be assembled by the end of the working hours.
Each planning problem is modelled in the standard language for classical/AI planning, namely the Planning Domain Definition Language (PDDL). The created modeled is then passed for solving to state-of-the-art AI task planning solvers that generate a plan. The plan contains the actions that must be carried out, as well as further pieces of information, such as the agent that is responsible for that action and the required tool. At last, the planned actions must be sent for execution to the agents, the robot and the human, through corresponding interfaces.
AI task planning applied to Sharework project
RWTH has modelled planning problems in PDDL for a general manufacturing scenario derived from the four Sharework project use-cases. RWTH has used this model within an extended version of the ROSPlan framework. ROSPlan is a framework that integrates AI Planning in the ROS middleware, sustains the planning process and the dispatching of the planned action for execution. The RWTH modelling and some of the extensions to ROSPlan are presented the following scientific paper:
- Improved AI Planning for Cooperating Teams of Humans and Robots (Stefan-Octavian Bezrucav, & Burkhard Corves). Proceedings of the Planning and Robotics Workshop of the 30th International Conference on Automated Planning and Scheduling (PlanRob).
- Case Study: AI Task Planning Setup for an Industrial Scenario with Mobile Manipulators (Stefan-Octavian Bezrucav, Malte Kaiser & Burkhard Corves). Proceedings of the Scheduling and Planning Applications woRKshop at the 31st International Conference on Automated Planning and Scheduling (SPARK).
- An Action Interface Manager for ROSPlan (Stefan-Octavian Bezrucav, Gerard Canal, Michael Cashmore & Burkhard Corves). Proceedings of the Planning and Robotics of the 31st International Conference on Automated Planning and Scheduling (PlanRob)
RWTH has deployed the planning models and the adapted ROSPlan framework in a complex Gazebo simulation. The simulation contains a representation of the general scenario, with one human and one robotic mobile manipulator cooperating in different assembly processes. The actions of the two agents are determined automatically and the execution is carried out autonomously.
With the deployment of AI task planning methods and special modelled planning problems, RWTH shows that an autonomous execution of manufacturing processes by a mixed team of mobile agents (robots and humans) is possible in simulation. In the next steps, the validation of this approach on real uses-cases will demonstrate that a real cooperation between humans and robots becomes reality.

Stefan-Octavian Bezrucav
Stefan-Octavian Bezrucav is pursuing his Ph.D. at the Institute of Mechanism Theory, Machine Dynamics and Robotics (IGMR) of RWTH Aachen University. He holds a bachelor’s and master’s degree in Computation Engineering Science. His main research area is the automated task planning for mixed teams of mobile manipulators, humans and robots, spatial reasoning, and complex simulations.
IGMR Team in Sharework
Stefan-Octavian Bezrucav (Task Planning, Reasoning), Nils Mandischer (Cognition, Safety)