Artificial intelligence in robotics brings huge potential for manufacturers ready to embrace it.
Companies are in a race to embrace new technologies to drive digital transformation and enable Industry 4.0. Intelligent adoption of artificial intelligence (AI) and machine learning to improve automated processes can be hugely rewarding. According to a review by BCG and MIT Sloan Management, nearly 85% of executives expect AI to enable them to obtain or sustain a competitive advantage. By allowing machine learning algorithms to utilize Internet of Things (IoT), sensor data and images, configurations can be implemented in real-time to further optimize production facilities and reduce downtime to a minimum. This can also improve quality, reduce cost and accelerate time to market. However, only approximately 9% of manufacturing organizations are leveraging artificial intelligence today. Why so few? Likely because it can be challenging to know where to focus and how to get started.
The role of artificial intelligence in robotics
As you consider how AI can improve automated processes, you’ll quickly realize that this largely translates into understanding where and how to implement artificial intelligence and machine learning in robotics. In this context, automation and robotics set the precedent and groundwork for using AI in your broader smart factory initiatives. To identify opportunities, we can categorize the applications of robotics according to the following:
- Product Assembly – One of the most common applications of robotics in contract manufacturing is in the assembly of products. Robotic manipulation of components in this environment is difficult. However, when paired with computer vision systems and products designed for automation, accurate distinctions can be made with greater consistency and efficiency. Examples here would be Printed Circuit Board Assembly (PCBA) or higher level assemblies, such as enclosures or single-use medical devices.
- Production Processes – Leveraging robotics for a single process across multiple products is another great use for industrial robotics. Processes in the electronic manufacturing industry such as bonding, fastening, welding, soldering, spraying and masking help to create repeatable actions, which are helpful in controlling key process or performance indicators.
- Assembly and Process Inspection – Using computer vision with a robot can provide inspection for robotic assembly or robotic process to ensure the value intended to be added by the robot has indeed been added. By using a robot to move around a camera, the limitations of 2D inspection can be mitigated by taking images at multiple angles with multiple lighting applications. Furthermore, advanced 3D systems metrology and digital inspection tools significantly increase product and manufacturing quality, rapidly identify process issues and build productivity.
- Machine Tending and Packaging – Automation is an everyday element of the packaging industry. From pick & place to automated palletizing, robotic systems are making packaging faster, flexible and easier. Performing mundane, repetitive tasks for workers they improve product quality and worker safety.
Where to start with machine learning in robotics: Automated Optical Inspection (AOI) for PCBA
By focusing on applying machine learning in robotics to one application at a time, manufacturers can fast-forward the implementation of new technologies and use it as the starting point to evaluate potential further investments.
Adopting AI in the manufacturing space can appear to lack practical feasibility. Sometimes that’s because we want to tackle too big a process all at once. A more achievable starting point can be to leverage AI algorithms in machine vision inspection, for example. This has become a way of incorporating the overall AI trend into 3D AOI. This key technique is part of the manufacture and test of electronics printed circuit boards (PCBs). In general, machine vision is used to inspect images before the boards enter the solder process. Those assembly line systems require skilled engineers to build and hard-code algorithms to help the system identify good and defective components — each and every time a board is introduced to a production line. In addition, the algorithm programming has limitations and is only as robust as the engineer is talented at creating it.
While effective, these systems usually lack the ability to learn or adapt. Still today, algorithms frequently mistake good assemblies with bad assemblies and deliver "false-calls", which impact the efficiency of overall production. A high false-call rate results in “humans in the loop” to check component diagnoses, which can amount to hours each day.
Today, manufacturers are faced with new AOI challenges:
- False Call Reduction - Many false calls are due to inappropriate programs (un-optimized), partly due to inexperienced programmers or human error. Auto-programming is expected to overcome such programming challenges by setting optimization thresholds.
- Escape Reduction - While false calls are not actually defective, escapes are defects that are not indicated. In a good AOI system, both of these should be low. AOI escapes can be divided into machine escapes and review escapes. Machine escapes, a direct escape from a machine, can be minimized with auto-programming. Review escapes, an escape from a review operator during image buyoff, can be minimized with auto-buyoff.
- Programming Time Reduction - Auto-programming can help speed up AOI programming time for NPI (New Product Introduction). Machine learning algorithms can replace the original engineered algorithms and are able to learn the nuances and parameters that define a “good” product, thus reducing the number of false calls and identifying defects to prevent escapes.
It is recommended that an integrated AI library is built internally via machine learning. This library generates the output of threshold values, inspection requirements and the pass/fail criteria for various types of component packages and enables AOI auto-programming and AOI auto-buyoff. Once at scale on all inspection machines, the data, or training set, will continue to grow and the algorithm will gain accuracy. The continuously compounding dataset will train the machine learning algorithms and continue to produce better results.
Understanding emerging algorithms for machine learning in robotics.
As AI and automation continue to advance, two categories of machine learning algorithms are important to understand:
- Many vision and camera inspection providers are developing and providing teachable algorithms with their off-the-shelf inspection systems. These algorithms allow specific parameters within the algorithm to be defined and then learned as high quantities of parts and assemblies are inspected. It is also advantageous to put known-bad cases through the process to decrease the learning time.
- For custom algorithm development, a specific set of parameters with complex characteristics need to be defined and quantified before the algorithm can be effective. Each algorithm is unique to that assembly, but a base process can be used to ensure the development of that algorithm is robust and reliable. Significant concept and verification testing are required to ensure efficiencies and effectiveness on the production floor.
Building the framework to support machine learning in robotics
Let’s stay focused on automated optical inspection as a starting point. What kind of structure and resources will you need to support it? In order to ensure that this process is valid for production, it’s necessary to store very large images from the production floor to create the database that will eventually be a training set. Today, in most factories, most of the implemented databases simply do not have the capability to adequately store such large amounts of data. A team of database analysts, control engineers and machine learning experts have to work closely together on a proof of concept database and storage hardware infrastructure. Furthermore, manufacturers of security-critical applications such as aerospace, defense and healthcare have to weigh issues with compliance and data security and protection.
The validation process surrounding these platforms when upgrading to machine learning-based algorithms has some risk as well. Manufacturers have to determine the best method for recreating the IQ, OQ, PQ (Installation Qualification, Operational Qualification and Performance Qualification) processes. Industrial standards around machine learning will need to be investigated to ensure that the customer’s regulatory compliance protocols are all met. The value of machine learning diminishes if the system cannot be validated effectively.
Machine learning in robotics: start small and build
Focusing on one application area to roll out a production-wide AI initiative can pay off in the long run. The automation and robotics team is able to collect information and scope programs for additional robotic systems to benefit from similar algorithm development. Leveraging the success of each narrow machine intelligence victory should result in a more robust production solution, not only for assembly and process inspection, but also across all four of the previously mentioned categories.
We have yet to reach the full potential of robotics and machine learning, but current applications of artificial intelligence in robotics are promising. The primary lesson is this: Start small. Build from there. As you learn from your own application of AI in robotics, you’ll be able to incrementally improve more and more processes. From there, the sky’s the limit.