Automated handling of flexible parts

Flexible components such as cables, hoses or textiles pose a major challenge for industrial automation due to their complex dynamics and high shape variability. In production environments, many process steps are therefore still carried out manually. Existing automation solutions are usually rigidly geared to the specific application, require special devices and require complex adaptation in the event of product changes. Current research work at the WZL at RWTH Aachen University uses the example of robot-based handling of T-shirts to demonstrate the use of modern AI approaches.

A frequently considered AI-based alternative to rigid automation solutions is reinforcement learning. During training, the robot system receives a reward for each of its actions based on an initially defined reward function and thus learns to perform a task. In practice, however, reinforcement learning approaches often fail to define a suitable reward function because the target state is difficult to quantify, particularly in the case of textiles.

Imitation learning does not require a reward function and thus avoids the difficulties of quantifying the goal. Instead, it uses the intuitive knowledge of humans and learns directly from expert demonstrations. These demonstrations are often recorded by manual teleoperation of the robot. The AI system receives the actions from the demonstration (the robot movement) together with the observation (the camera images) and learns to independently determine the appropriate action for new observations.

Structure at the WZL

For the experimental investigation of the imitation learning approach, a two-armed robot system was set up at the WZL in cooperation with an industrial partner, on which tasks from fabric and laundry handling are implemented. The test platform consists of two Franka FR3 robots that are controlled via the Robot Operating System (ROS). The teleoperation of the robots is implemented by Meta-Quest 3 VR controllers. The person controlling the robots can move the controllers in space and the robots imitate the movement in real time in Cartesian space. This enables intuitive and precise teleoperation.

The demonstration setup is supplemented by a camera-based sensor system that enables comprehensive visual recording of the scene. A top-down camera captures the global action and material spread, while two wrist cameras on the robot arms record the interaction up close. This enables both context understanding and fine-grained motion capture, which is crucial for the training quality of modern vision-language-action models.