How robots can help us understand reinforcement learning

A robotic arm donated by Kindred AI is giving students a new way to study reinforcement learning at UAlberta.

Katie Willis - 12 November 2019

A robotic arm, donated by Kindred AI, is changing the way that students learn about reinforcement learning (RL) in the University of Alberta's Department of Computing Science. And this new approach has the potential to create innovative solutions to real-world problems.

"Teaching robots to learn how to solve complex tasks in unstructured environments is one of the grand goals of RL," explained Rupam Mahmood, assistant professor and former Kindred AI research lead. "Imagine a fully autonomous AI system that explores the world, adapts to changing environments, and solves increasingly complex tasks in a self-motivated way. This type of system could navigate and work in many modern environments-such as filling orders in a warehouse."

Prior to joining UAlberta in 2019, Mahmood worked at Kindred AI, a Canadian-based company creating robotic solutions for e-commerce problems. Kindred AI combines RL systems and robots to improve the efficiency and accuracy of tasks like warehouse and fulfillment sorting. The technology rivals a human's ability to identify, pick, and place items, and has the potential to be applied in other industries.

Return to campus

After demonstrating the great potential of real-world RL at Kindred AI, Mahmood sought to pursue the future of RL research, re-entering the world of academia.

"Kindred sees how academic research is fundamentally important for the future of the robotics industry," said Mahmood, who is a principal investigator in the Reinforcement Learning and Artificial Intelligence (RLAI). "Understanding the need for sophisticated physical robots for my research program, they donated one of their industrial robotic arms to the Department of Computing Science when I joined UAlberta to pursue my long-term research goals."

In offering the robot arm, James Bergstra, cofounder of Kindred and head of AI research said, "We hope our gift of robotic hardware to the University of Alberta inspires students and researchers to bring their ideas to life in the physical world. Recent advances in the use of neural networks for supervised and unsupervised learning show that with enough real-world data, neural networks can recognize patterns in images, sound, and text. We hope this gift helps encourage reinforcement learning research beyond simulation."

This fall, Mahmood is teaching a graduate-level course called Reinforcement Learning with Robots, designed to prepare students for RL research, which will readily carry forward to real-world problems.

"In my research, I will use the arm to make progress in both theory and applications of RL. My lab will investigate whether current RL methods applied to real robots can adapt to new environments and accomplish tasks in a self-motivated way by creating many different tasks based on this robotic arm."

In addition to his role as an assistant professor in the Department of Computing Science, Mahmood is a fellow at the Alberta Machine Intelligence Institute (AMII). Drawing from world-leading academic research at UAlberta and other institutions, AMII helps Alberta workers reskill and upskill for high-demand careers in artificial intelligence, and guides Alberta-based businesses as they implement artificial intelligence across operations and build their in-house capabilities and teams.