Using Math and Logic to Train Robots
Using math to teach robots how to think isn鈥檛 science fiction. It鈥檚 Brendon Johnson鈥檚 research.
A student in the 黑料老司机鈥檚 Master of Science in Computer Engineering program, Johnson spends his days in the Bio-Inspired Robotics Lab under the guidance of Professor Alfredo Weitzenfeld. There, he is doing more than simply programming machines. He鈥檚 teaching them to learn.
鈥淚鈥檝e always loved math and logic,鈥� he said. 鈥淲ith reinforcement learning, you take math and logic and let the machine figure things out over time. It becomes more than just programming. It becomes learning.鈥�
He shares an example of how this works in the lab. In one project, a small robot on wheels edges forward, hesitates at a junction, turns left, bumps into a wall, and tries again. It鈥檚 not following a map or a script. It鈥檚 learning, slowly building knowledge from each success or mistake.
鈥淩einforcement learning mimics how animals or humans learn,鈥� Johnson explained. 鈥淪o, if you do something right, you get a reward. If you do something wrong, you get punished. The idea is to train a robot or agent to figure out the best action for each situation.鈥�
So rather than feeding the robot a rigid list of instructions, Johnson designs algorithms that allow it to experiment. The machine earns rewards for behaviors that help it navigate a maze or complete a task and learns to avoid the ones that don鈥檛. Over time, it becomes smarter and more capable.
How do you reward a machine?
鈥淯sually it鈥檚 a numerical score,鈥� Johnson explained. 鈥淭he robot gets a higher value when it performs a desired action, like reaching the goal, and a lower or negative value when it fails. Over time, the machine learns which actions earn better scores.鈥�
A Program That Builds Deep Knowledge and Practical Skills
After earning a bachelor鈥檚 degree in computer engineering, Johnson knew that he wanted to specialize in reinforcement learning, a field he described as 鈥渒ind of niche.鈥�
鈥淚 graduated from a great program at Brigham Young University, where they had great engineers and computer scientists, but none of them had conducted research in this area. I learned about the work Dr. Weitzenfeld does in his lab and that was a big part of the reason I chose USF,鈥� he said. 鈥淚 liked the mix of robotics and reinforcement learning which is inspired biology.鈥�
The New Mexico resident said the affordable tuition helped with his decision, too. 鈥淚 was looking around the country for good schools, discovered the robotics program and research opportunities, and realized USF was a good value, too,鈥� he said.
Johnson opted for the thesis track of the MS in Computer Engineering program. It gave him a strong foundation in computing systems, and he had the chance to take varied electives while specializing in machine learning and robotics.
He says the variety of elective courses made the program enjoyable. He is quick to point out, though, that interesting and enjoyable doesn鈥檛 equate to easy. The courses were rigorous, and the faculty pushed him at times.
鈥淭he courses were challenging but always applicable,鈥� he said. 鈥淲hat you learn in class, you can test in the lab. That combination of theory and practice is what makes it stick.鈥�
Faculty support also played a central role. Under Weitzenfeld鈥檚 mentorship, Johnson gained the confidence to refine his ideas and take initiative as a researcher.
鈥淚 had the freedom to explore and solve problems on my own, but I always knew I could get feedback or guidance,鈥� he said. 鈥淭hat balance helped me grow.鈥�
Looking Ahead
Johnson defended his thesis last month and will soon start the doctoral program at USF. In his paper, 鈥淗ierarchical Reinforcement Learning in Multi-Goal Spatial Navigation with Autonomous Robots,鈥� he writes about testing manual versus automatic sub鈥慻oal creation, examining termination function frequency and demonstrating clear advantages in performance and adaptability. In the doctoral program, he plans to explore how experience-driven learning and computer vision can work together to help machines better interpret their surroundings.
鈥淚t鈥檚 exciting to build something that learns,鈥� he said. 鈥淚t鈥檚 not just programming a machine. It鈥檚 teaching it how to improve, adapt and solve new problems, and that鈥檚 what the future is about.
Looking ahead, Johnson hopes to apply his research in real-world tech fields, especially autonomous systems. 鈥業鈥檇 like to do reinforcement learning in robotics,鈥� he said. Whether in cars, drones, or future smart devices, he鈥檚 focused on what鈥檚 next: Building systems that can truly learn.