Artificial intelligence (AI) applications improve the ability of industrial robots to react quickly and precisely to sensor and image data. Integrating AI functionality directly into the robot control system enables not only accelerated object recognition, but also increases in quality, reliability and speed, as well as extending the possibilities for working with people.
Robots take over the jobs that are too dangerous, too complicated, too heavy, too dirty or too monotonous for people. In industry, automation of such tasks by means of robots has become standard. There have always been attempts to make the mechanically very flexible machines react flexibly to changing situations outside of rigid programming. For example, the freely programmable movements and gripping operations are often not rigidly specified, but rather react to the input of sensors.
Robots need greater flexibility
Despite needing flexibility, fixed, algorithmic programming is always involved, which is usually optimised for high working speeds. In most applications, the supplied robot controller executes the programs. In particularly time-critical cases, for example for the demoulding of new plastic parts in injection moulding machines, the level of integration is greater. In this area, the machine control and drive control units for the individual axes directly control the robot kinematics.
In many cases, robots cannot and are not intended to replace humans, but rather aid and support them. The idea is for people and machines to work closely together in the workplace in a common process. The vision extends further to robots designed to support people in hospitals and care facilities, but also in their own homes or on their daily commutes.
Collaborative robots, or cobots, represent the first step in this direction. However, these require sophisticated additional sensors, for example in the form of a touch-sensitive skin, as well as further adjustments in order to enable this cooperation without compromising on safety.
Above all, the unpredictability of human actions does not make it any easier to set the robots free from their cages. This is why so far, they are only able to work collaboratively at greatly reduced speed. Their use outside an industrial environment fails not least due to the unknown and changing environment that no software developer can anticipate and take into account.
Sensory perception for robots
In robotics, “bin picking” to remove workpieces delivered in bulk and working safely hand in hand with people are considered to be classic challenges. Solutions for both cases result in connection with image processing systems and highly developed sensors. These include 360° laser scanners, time-of-flight (TOF) cameras or stereo 3D depth cameras.
They not only give the robots the opportunity to recognise the objects to be gripped, but also to perceive the environment in three dimensions. In addition, it is also possible to carry out a visual quality control of the gripped parts during the process, allowing the robot to remove parts found to be of poor quality at this stage. This undoubtedly results in lower costs than later in the overall process.
Independence through AI
Although such tasks can also be solved purely algorithmically, it takes more to turn robots into both cooperative and efficient colleagues. They need the ability to adapt to new situations. Artificial intelligence applications such as deep learning are very well suited for recognising patterns in image data. Their real-time evaluations enable a significant improvement in accuracy in detecting objects and protecting people in collaborative use. At the same time, they can be used for quality assurance. In addition, machine learning methods can be used to provide the robot with a statistical time advantage through predictive positioning.
Let’s take the example of an assembly hall at a tyre dealer. A stationary six-axis articulated robot on each side of the lifting platform dismantles the wheels. Using a 3D vision system, it finds out where the wheel is located in the room, whether and if so which hubcap needs to be removed and how many screws need to be loosened at which exact positions. If one of them is missing or is visibly damaged, it orders a new one in the spare parts warehouse. Before the robot puts down a wheel, it can also determine and register the tread depth so that it becomes visible in the storage file or on the customer invoice.
The wheels to be assembled are fetched by an autonomous mobile robot (AMR). As all three machines are constantly observing their surroundings—the AMR also uses this information to navigate—collisions are avoided. The assembly robots can also move into position as the AMR approaches so as not to lose any time when removing the wheels. They identify the exact location of the screw holes before removing them. Before assembly, they let each wheel rotate to detect and document any imbalance. After successful assembly, they load the removed wheels onto the AMR again.
Using these robots as described in the example would remove a considerable burden from employees, who would have to perform fewer physically strenuous activities such as mounting the tyres as well as less monotonous work such as screwing, enabling them to focus instead on other tasks.
How does AI reach the robot?
AI applications are available in large numbers as software-as-a-service (SaaS) in the cloud. However, many users have concerns about the reliability of data connections given the considerable volumes of data involved. These can be a significant cost factor. It is therefore obvious to move the inference calculations to the edge of the system.
Whereas AI applications previously had to be outsourced to expensive, high-performance systems, the availability of specialised AI coprocessors and their integration into single-board computers or control computers now enables the extremely computationally intensive AI operations to be carried out directly at the point where they are required.
There are already robots with correspondingly powerful graphics processor boards, which are also suitable for processing some AI tasks. In addition, more and more suppliers are bringing developer boards and single-board computers with integrated AI accelerator chips (tensor processing units, TPU) to market. Connected to the robot controller—the appropriate connections are already provided on some models—or integrated into the image processing systems, these enable machine learning and deep learning applications to be executed directly on the robot.
Teaching robots how to learn
AI cannot give robots the cognitive abilities of humans. It allows them to adapt their behaviour to changing environmental conditions in an unstructured world on the basis of statistical empirical values based on the information gathered. Therefore, it would not be possible to realise the scenario described above using classic programming, or if it were, it would require an enormous amount of programming work, particularly if this also included cooperation with people.
However, cognitive robots, i.e. robots equipped with AI, have to learn this behaviour first by identifying recurring patterns, laws or anomalies on the basis of neural networks. It is best to not do this for the first time in the system in which failed attempts would have a negative impact on efficiency and other aspects, but beforehand. For this purpose, the system integrator can have the digital twin of the system run training cycles in the computer model in a short time without posing any risk. However, pre-trained inference models can also be used for this purpose and are increasingly becoming part of the standard equipment of AI hardware and software. In addition, it is a good idea to limit the degrees of freedom to that which is necessary.
Expert knowledge is needed for this purpose. This also means that AI does not make software developers and control programmers superfluous. It enables them to take a different approach to problems and gives them different, often more convenient tools. However, learning to use these does take some getting used to and practice.
Luckily, not all industrial robots are large, expensive and complex to install. reichelt elektronik GmbH & Co. KG offers a wide range of low-cost robots that are ideally suited as a development platform, for laboratory applications as well as for training and experimentation. These include autonomous, four-legged robots that are reminiscent of dogs.
The devices are freely programmable via an open robot operating system. The middleware has a rich library of ready-made functions to make getting started easy. The electronics and computer technology distributor based in North Germany can also supply the appropriate sensors and camera systems as well as the corresponding accessories and AI-compatible single board computers from a single source.
Smarter robots, that is, those with AI on board, can do more. They boast greater independence, higher quality, reliability and speed, and are better suited to working with people. This makes them economically viable even in smaller applications. Their future may only just have begun, but it is clearly picking up speed and there seems to be no way around it. Companies that prepare for it now will be at a clear advantage.
Images: Adobe Stock
