Using reinforcement learning for efficient trajectory planning of parallel robots in handling flexible objects

• Applicants: Professor Dr. Tobias Glasmachers; Professor Dr.-Ing. Bernd Kuhlenkötter
• Subject Area: Automation, Mechatronics, Control Systems, Intelligent Technical Systems, Robotics ; Production Automation and Assembly Technology
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 555175075

Parallel robots have excellent dynamic properties due to their kinematic structure and lightweight design, which are reflected particularly in high speeds and accelerations. They are typically used in sorting tasks, such as in the food industry. However, the problem with these robots is that fluctuations in production can have a very strong impact on the dynamics of the system due to their lightweight design. Previous control methods for parallel robots also have shortcomings, so that the high dynamic properties and potentials of the robots cannot be fully utilized. As a result, the achieved speeds and accelerations in practice are often significantly lower than those specified in the data sheet. In addition, unwanted vibrations of the handled object can occur, which can lead to damage to the object. To increase the efficiency and optimized movement execution of highly dynamic parallel robots, novel planning and control concepts are required, which should be developed and implemented using methods of machine learning. The central idea is to use reinforcement learning as an approach for trajectory optimization of parallel robots. Early planning and optimization of the robot trajectory brings significant advantages, as the process and cycle time can be minimized by optimizing the robot trajectory, and unwanted vibration phenomena of the handling object can be reduced by suitable vibration compensation in advance. To implement reinforcement learning, a large amount of data is required to create a suitable learning database for trajectory optimization. However, a database based on measurement data requires a very large number of experimental series with the real robot system, which is then not available for production. In addition, new experimental runs are necessary whenever the smallest changes occur in production. To minimize the required experimental effort, a suitable simulation environment of the robot motion behavior under varying production conditions is to be implemented. A sub-problem is the calculation and simulation of robot movements in which the object handled by the robot is deformable, since deformability is very difficult to simulate in detail. The goal is to develop a suitable simulation environment for parallel robots, which realistically represents the movements of the robot, its different loads, and the objects that are handled. The simulation environment then serves as a data provider for the methods of reinforcement learning and trajectory optimization to be developed.

• DFG Programme: Research Grants