Towards the automation of woven fabric draping via reinforcement learning and Extended Position Based Dynamics

Abstract

The draping process in the preforming stage of composite manufacturing is very cost- and time-expensive and requires substantial manual labor. One strategy towards automation is the use of collaborative robots. Recent advances in AI have made it possible to train robots on difficult real-world tasks with reinforcement learning. However, training a robot using reinforcement learning is practically challenging and leveraging simulation is often the only option to use reinforcement learning in real-world settings at all. Existing FE models, which are commonly used for optimization of preforming processes, are too slow for reinforcement learning training. We addressed this issue by developing an XPBD-based surrogate model, drastically reducing simulation times compared to a classic FE model. With the achieved speedup, the training of a reinforcement learning agent became feasible and a draping trajectory could successfully be transferred to a real-world cobot, demonstrating the potential and capabilities of this innovative approach.