ScotEng Blog: AI-Powered Control Systems: Revolutionizing Engineering Performance
Author
Oliver King-Smith, CEO of smartR AI
5 minute read
While much attention has been focused on AI’s capabilities in data analysis and decision-making, its potential to revolutionize control systems and mechanical performance remains relatively unexplored. Recent innovations demonstrate how AI can transform basic hardware into sophisticated, adaptive systems that outperform traditional control methods.
Reimagining Control Theory with AI
Traditional control systems often rely on precise hardware and complex mathematical models. However, by leveraging AI, engineers can now achieve remarkable performance from relatively simple hardware. This approach shifts the paradigm from rigid programming to goal-oriented learning, where systems adapt and optimize based on real-world data.
From Simple Hardware to Sophisticated Performance
Consider a fascinating experiment with a low-cost robotic arm tasked with continuously bouncing a ping pong ball.
We used a toy £200 robot arm with imprecise feedback that wasn’t rigid when operated. Rather than programming explicit movement patterns, the system we call SERLE was given specific goals and allowed to learn from actual performance data. Interestingly, when initially given just two performance goals, the AI discovered an unexpected solution – simply balancing the ball on the paddle while moving up and down. Adding a third goal constraint led to the desired dynamic bouncing behaviour, highlighting how AI systems can find creative solutions within given parameters.
The SERLE engine we created is based on a Sample Efficient Reinforcement Learning algorithm. There were actually two different machine learning components. One of them was the Policy network which took in the location of the ball (as determined from an inexpensive Intel RealSense camera) and the reported angle of the joints.
The second network call the Uncertain Dynamics Engine modeled the real world from all the data that was collected. The Uncertain Dynamics Engine actually understands how confident it is when it predicts the real world. The two models worked together to explore the space of solutions. When the policy network thought it had a solution, we would run in the real world again, collect more data, and then update the Uncertain Dynamics Engine to better understand the real world.
This collaborative approach to learning speeds up how quickly you can solve problems, and massively reduces the amount of data you need to collect (in this case we needed less than 1% of the data we would have needed if we just tried to train the policy network in the real world).
Enhancing Traditional Control Systems
AI’s impact extends beyond novel applications to improving conventional control methods.
In one compelling case study, AI was used to enhance a PID controller for an optical system that was facing stability issues. During manufacturing just over 10% of the units failed. It was expensive and hard to correct the manufacturing problem, because you saw the results only after the whole unit was assembled.
Rather than pursuing an expensive hardware redesign, the team implemented AI to fine-tune the PID parameters and introduce non-linear corrections.
This innovative approach salvaged existing hardware while achieving the required performance specifications.
Rapid Ice Cream Manufacturing: A Sweet Success Story
Perhaps one of the most delicious applications of AI in control systems comes from an experimental ice cream machine project. The challenge was to create perfect ice cream in just 90 seconds using inexpensive motors. The system needed to maintain precise blade speeds while advancing at exactly 0.012mm per revolution through frozen material – a task complicated by varying fat content, temperature fluctuations, and non-linear blade contact forces.
Instead of relying on expensive high-precision motors and gearboxes, the smartR AI team developed an AI-controlled system using basic components: a current sensor for the motor and a simple rotation counter. The AI learned to compensate for variations in material properties and mechanical interactions, delivering consistent results from modest hardware. The only downside? Too much temptation for taste-testing during development.
Building Robustness Through Learning
These examples highlight a crucial advantage of AI-enhanced control systems: their ability to adapt to real-world conditions and variations that might be impractical to model mathematically. By learning from actual performance data, these systems can achieve robust operation despite using simpler, more cost-effective hardware.
One challenge engineering teams have with AI is the leap of faith that they can underdesign rather than overdesign. Almost always engineers select components that exceed their requirements. Only when a component is not available, or way too expensive, do engineers look at components that are more cost effective. With AI, the universe of parts that you could select opens. Effectively AI can supercharge the capabilities of lower priced components.
For example, in the robot arm example, the robot angle sensors were poor, and the servos wouldn’t return to the exact same location. Yet, the AI could overcome this by learning the quirks of the system.
This new type of engineering does require thinking about problems differently. AI offers the ability to model complex relationships, noise, and time delays, but the specs are not printed on the box. You need to develop a sense of what AI can achieve, and then rapidly prototype to prove the results.
The benefits are machines that perform better, operate over wider extremes, age more gracefully, and perhaps most importantly cost less money to make.
Looking Forward
The integration of AI into control systems represents an exciting frontier in engineering. As these techniques mature, we can expect to see more applications where intelligent control enables sophisticated performance from simple hardware. This approach not only reduces costs but also opens new possibilities for adaptive, resilient systems across various industries.
The potential for AI in mechanical and electrical control systems remains largely untapped. As more engineers explore this space, we’re likely to see increasingly creative applications that challenge our traditional approaches to system design and control. Connect with us to join a community of innovators who are pushing the boundaries of what’s possible with AI-enhanced control systems.
At smartR AI, we’re passionate about delivering cutting-edge AI solutions. If you’re looking to explore the potential of AI in your operations and harness the power of AI to drive innovation and growth, don’t hesitate to contact us at smartR AI to discover how our expertise can help drive your success.
About Oliver King-Smith
Oliver holds a PhD in Mathematics from UC Berkeley and an executive MBA from Stanford, and is an innovator with expertise in Data Visualization, Statistics, Machine Vision, Robotics, and AI. As a serial entrepreneur, he has founded three companies and contributed to two successful exits. At his latest company, smartR AI, Oliver King-Smith spearheads innovative patent applications harnessing AI for societal impact, including advancements in health tracking, support for vulnerable populations, and resource optimization. Throughout his career, Oliver has been dedicated to developing cutting-edge technology to address challenges, and today smartR AI is committed to providing safe AI programs within your own secure and private ecosystems.
LinkedIn profile: https://www.linkedin.com/in/oliverkingsmith/
Email: oliverks@smartr.ai
About smartR AI
We believe that data and AI play a key part in a sustainable world. Our mission is to facilitate and empower organizations to extract real value from their data in an ethical, responsible, and sustainable manner using cutting edge AI technology. We provide solutions that are risk-free, controllable, and financially viable for any small or medium-sized enterprise, especially those who are innovators and challengers wishing to shape the future.
We like to think of ourselves as the human face of generative AI, in a world where today’s technology is causing confusion. So, every company can now build its own private AI smarter.
