The AI-PRISM project is oriented towards a human-centred AI-based solutions ecosystem targeted to manufacturing scenarios with tasks difficult to automate and where speed and versatility are essential. To facilitate the assessment of the performance, transferability, scalability, and large-scale deployment of AI-PRISM solutions, we will conduct research under real operational environments in five pilots involving key manufacturing sectors: furniture, food/beverage, built-in appliances, Electronics and one generic discrete manufacturing.
Regarding the discrete manufacturing demonstrator, researchers from Robotnik, Cranfield University, Vigo Photonics, Profactor, Fondazione Bruno Kessler and WINGS visited KEBA Group facilities in Linz, Austria. Combined with the growth of the past years, this company has bundled its business into three strong areas for the future: industrial automation, handover automation and energy automation. The first focuses on OEM (Original Equipment Manufacturers), the second is a project business producing self-service machines in the banking and logistics sector, and the third manufactures wall boxes for e-cars and heating control systems.
This use case focuses on deploying future robotic systems on the shop floor where untrained users coordinate human and mechanical tasks more naturally. In other words, we aim to implement and demonstrate AI-Control for natural, multimodal human-robot collaboration. Before AI-PRISM, a new prototype of an AI-based control system for future robotic automation exists, providing architecture with interfaces to ROS and relevant machine learning libraries. It is an initial technical basis to improve the speed and flexibility of (robotic) control on the shop floor. However, the solution needs more support for the multimodal training of robots for complex and collaborative shop floor processes.
Collecting data at the PBC testing area and interviewing operators
During their visit, the tasks analyzed by AI-PRISM researchers were the recognition of PCB shapes, grasping processes for a given PBC, manipulation correction and bin packing or deposition of PCB. Moreover, all the tasks mentioned above aim to be configurated in an interactive process workflow (using an easy-to-use interface) by the user.
The first task involves recognizing the PCB either using model-based methods (CAD, etc.) or model-less approaches (AI-based vision). The second process is based on AI-enabled grasping (state-of-the-art) approaches combined with user configuration to deal with variants and a quick teaching process. possible) and feedback from the user after the process should help the process. After grasping the object (PCB), the object’s alignment in the robot’s ‘hand’ might not be suitable for the next process step (testing, deposition). So a correction step to re-orient the robot ‘hand’ is required. Then, the bin packing’s goal is to efficiently place the tested PCBs on a plate in the transport box.
Through this first data collection, researchers observed and interviewed two operators, one novice and another experienced. They found that the novice user compared to the experienced user, spent more time fixating on all AOIs, but mainly on the monitors also the board test area. While the experienced operator had hardly any fixation on the scanner location, the novice user spent around 5% of fixation time in this AOI.
Alleviating execution failures
One of the challenges when configuring the tasks in a process flow is the cases where execution fails, and recovery strategies should be applied to alleviate the failure. For example, whenever the operator observes a fault, they try to figure out the reason – it could be the board, or the machine needs fixing. During the observation by researchers, the testing failed five times, and the operator managed to fix the failed process three times.
In the discussions with the operators, they realized that the most error-prone parts of the process were placing and removing the board and scanning. This is because the method relies on procedural knowledge and inspection. Yet, the self-efficacy of fixing some fails and independence in the process were important aspects contributing to operator satisfaction and the need to remain with the development of AI-PRISM technologies.
For all the tasks, the aspect of easy reconfiguration or programming of tasks by workers with a low level of expertise in robotics and control, remote monitoring of certain operations, machinery control and checking of potential impairments would be significant in this demonstrator. The AI-PRISM team is excited to see where this project will take us. We will apply the same methods and technologies in the other industries that conform to our five pilot use cases, requiring high manual dexterity and expertise.
Do you want to learn more about AI-PRISM and the team behind the project? Subscribe to our newsletter and follow us on LinkedIn and Twitter, so you don’t miss a thing!