Min Reading Time
Quality control made easy: Thanks to artificial intelligence, robots can support their human ZF colleagues more efficiently in production. Research and development on this topic is being performed in Saarbrücken.
Lars Weitbrecht,
author_image
Lars Weitbrecht originally comes from the music and gaming industry, but in addition to holding a game pad or guitar in his hand, he also enjoys the power of the pen and the feel of the steering wheel.
In many ways, the human being is superior to the computer. However, this is not the case when it comes to precisely performing monotonous tasks over a longer period of time. "Monotonous work that constantly requires full attention is a concentration killer. The human brain is not designed for this," explains Dr. Tobias Masiak, project manager in the Innovation & Technology sector of the AI Lab at the ZF Saarbrücken location. There, Masiak and his team develop the "Smart Camera Bot", which uses its new, specially developed AI algorithms to save its human ZF colleagues from just such a concentration killer.

At the Saarbrücken plant, around 10,000 transmissions of different types roll off the assembly line every day. Each part passes through several levels of quality assurance before it is delivered to the OEM customer. One of these levels is a visual check: Are all connections wired correctly? Are protective caps in the right place? Do these protective caps have the right colors? Is the paint intact?
This is exactly the kind of task in which a digital brain can do better than the human brain. "Our Smart Camera Bot can take over this monotonous task and thus optimally relieve the colleagues," says Masiak. The prototype consists of a small wheeled platform for maneuvering, a multi-articulated arm with a camera attached to it – and powerful algorithms, which enable the bot to perform the quality assurance tasks independently and automatically. For the "Smart Camera & Robot Agent" research project, the AI Lab recently received a grant from the Saarland State Chancellery and the EU's European Regional Development Fund.

"Level up" through artificial intelligence

"Level up" through artificial intelligence

Camera solutions in quality control are actually nothing new. However, in conjunction with conventional image processing, they reach their limits when it comes to complex characteristics. In order to keep up with the human eye – that is, to collect sufficient visual information for evaluation – it is currently necessary to have at least two, but usually ten or more cameras with which the product can be scanned. The software that performs this evaluation would have to be written separately for each product variant. In the end, this means that it would be very difficult to use such a device on more than one product line. And it would still be expensive: Currently, the costs for such a solution amount to several thousand euros.
The Smart Camera Bot is quite different. "Using AI for quality control – this may sound like breaking a fly on the wheel. However, it's actually the most efficient way to address this challenge," Masiak explains. Thanks to AI, a single, comparatively simple (and therefore cost-effective) camera is all the bot needs. "This camera is made so powerful by our algorithms that the robot needs far less raw image information for precise object recognition and analysis."
In doing this, the camera bot is supported by a Group-owned product: the AI-capable control unit ZF ProAI. ZF developed this central computer to enable autonomous driving functions. Thanks to its modular design, the high-performance board can also be configured for easier, stationary use in the camera bot.
"In the end, the bots are just small high-end PCs with multi-articulated arms," says Masiak. "We also explicitly decided against controlling the AI via a cloud connection. The robots will therefore keep running even if there are network failures."

Virtual training for the camera bot

Virtual training for the camera bot

Thanks to its algorithms, the "unintelligent" bot becomes a valuable support for operation. Team colleagues are relieved and can dedicate themselves to other tasks. Load peaks can be better managed. At the bottom line, this means more efficiency and time savings.
But how does the camera robot learn how to perform its task correctly? This is carried out via virtual training. "We send the AI through a series of simulations to push it step by step towards the correct workflow," explains Masiak. In this way, the machine practices approaching the assembly line, moving the camera arm around the product, detecting the transmission as a whole and in detail, identifying errors and much more. "While doing this, the bot receives constant feedback. If it does something wrong, the system will give negative feedback, and the AI will change its approach in the next run-through. If successful, the bot will receive positive feedback and will accurately repeat the step in the next simulation." The bot automatically repeats the task until it always fulfills it 100% correctly.

Added value for man and machine

Added value for man and machine

"The highlight, however, is that – thanks to AI – our robot is smart enough to transfer what has been learned," says Masiak. This means that a bot can also differentiate between and analyze different design variants of a transmission on the basis of a single data record, without generating a warning or be overtaxed by the task. This again saves time, which would otherwise be necessary to painstakingly reprogram the algorithms. And, in case of workload peaks on a particular assembly line, more camera bots from other production stations can be used for support.
Masiak sees even more potential in feeding the recorded image data into the Group-internal ZF cloud. "The more often our plants use the Smart Camera Bots, the bigger the information library that the bots can access. In short, the greater the amount of data, the better the analysis capability of each individual robot." And, since this is a ZF-internal solution, the know-how will remain within the company as well.
The research project is currently ongoing until the end of 2022. "Then we can be sure that our bot is running safely and reliably. Until then we can make plans on how to roll out this concept extensively throughout the Group," says Masiak.