Skip to Content

Reinforcement learning and optimization based path planning for thin-walled structures in wire arc additive manufacturing

This research presents a path planning framework, RLPlanner, which uses reinforcement learning and Sequential Least Squares Programming optimisation for automatic deposition path planning in wire arc additive manufacturing of thin-walled structures. Addressing the limitations of existing path planning strategies, this paper demonstrates the framework's adaptability to geometry variations and its ability to adjust welding parameters.


Type of Work: Scientific Publication

Main Author: Jan Petrik

Affiliation: ETH Zurich

Co-Authors: Markus Bambach

Date: 5th May 2023

Journal: Journal of Manufacturing Processes

Online: ScienceDirect

The methodology of the research centers around the development of the RLPlanner framework, which incorporates reinforcement learning, particularly Proximal Policy Optimization (PPO), for determining the optimal deposition paths without human intervention. By automating the preprocessing, parameterization, and optimization steps, the framework adeptly handles various geometries and adjusts process parameters like welding speed and wire feed rate.
Key findings from the application of RLPlanner to wire arc additive manufacturing demonstrate its potential to improve the adaptability and efficiency of deposition path planning through the integration of reinforcement learning. Experimental results demonstrate the framework's ability to generate optimal paths for different geometries with variable welding speeds and wire feed rates, thereby reducing the need for manual intervention, speeding up the overall process and potentially reducing the margin for error. The next plan is to integrate this path planning algorithm directly into the printing process, using a CMOS camera to attempt to control the welding torch in situ.


  • Image 1
    Experimental setup located in Advanced Manufacturing Laboratory at ETH Zurich, which was used to generate and subsequently digitize experimental data. Image A (left) visualizes ABB welding robot IRB 1600 with Fronius welding power source TPS 500i CMT with which weld beads were printed. Furthermore, image B (right) shows Atos scanner, which was utilized to scan the experimentally obtained data.
  • Visualisation of a trained reinforcement learning agent that has to find a deposition path for a double bead geometry with nine pockets as fast as possible while minimising the number of stopping points. The desired geometry is represented by white dots, while the already imaginary deposited material is shown in green. The movement of the robot where it did not weld is then shown in blue. Finally, you can see whether the robot is welding or not by looking at the indicator above the geometry, which shows Power ON or OFF.