The challenge of variable components

Industrial assembly operations have long relied on robotic assistance, with automated systems handling tasks like screwing, joining, and inserting. However, challenges arise when dealing with components that exhibit significant variations, making full automation impractical and costly. Additionally, the production of small batches often proves to be economically impractical. Furthermore, operations that involve complex contact interactions between robots and their environments, such as press or click fits, are difficult to program manually.

Imitation learning for flexible assembly automation

In pursuit of enhancing the flexibility of production lines, making small batch sizes economically feasible, and reducing the programming burden for machine operators, Work Package 4 within the HARTU project is delving into imitation learning approaches for contact-rich robotic assembly tasks. At the Robotics Innovation Center of the German Research Center for Artificial Intelligence (DFKI) in Bremen, Germany, a simulated assembly line for consumer goods has been established, focusing on electric shavers from Philips Consumer Lifestyle B.V. (PCL).

The task involves placing the front panel of an electric shaver on a rotating jig, which transports the components to a painting line – at the moment this task is currently performed manually at PCL. Data collection occurs through user demonstrations, where an operator guides the robot as it performs the task. However, the objective is not merely to replicate these demonstrations but to enable the robot to generalise its knowledge. This means the robot learns how to perform the task under varying conditions. For instance, the front panels of electric shavers may differ in size and shape, and the position of the rotating jig may vary with each operation. These unpredictable variables cannot be efficiently handled through traditional automation, especially when the components change frequently, sometimes multiple times per hour.

The HARTU project leverages modern machine learning techniques, including imitation learning and inverse reinforcement learning, coupled with powerful simulation tools like MuJoCo. This endeavor aims to make future manufacturing processes more adaptable and to automate assembly tasks that are still manually performed today.