A new artificial intelligence (AI) algorithm has been developed by engineers at Northwestern University specifically for smart robotics. This algorithm is designed to help robots learn complex skills rapidly and reliably, which could greatly enhance the practicality and safety of robots in various applications such as self-driving cars, delivery drones, household assistants, and automation.
Known as Maximum Diffusion Reinforcement Learning (MaxDiff RL), the algorithm is successful because it encourages robots to explore their environments as randomly as possible in order to gain a diverse set of experiences. This “designed randomness” improves the quality of data that robots collect about their surroundings, leading to faster and more efficient learning in simulated robots and improving their overall reliability and performance.
In tests against other AI platforms, simulated robots using Northwestern’s new algorithm consistently outperformed state-of-the-art models. These robots were able to learn new tasks and successfully perform them on the first attempt, showcasing the effectiveness of the new algorithm compared to slower trial-and-error learning methods.
The research will be published in the journal Nature Machine Intelligence on Thursday, May 2.
Thomas Berrueta, who led the study at Northwestern, explained that the new framework provides more reliable results compared to other AI frameworks. With this new algorithm, robots are expected to accurately perform tasks every time they are turned on, making it easier to interpret robot successes and failures in an increasingly AI-dependent world.
Berrueta, a Presidential Fellow and Ph.D. candidate in mechanical engineering at Northwestern’s McCormick School of Engineering, collaborated with robotics expert Todd Murphey, a professor of mechanical engineering at McCormick and Berrueta’s adviser, on the study. Allison Pinosky, also a Ph.D. candidate in Murphey’s lab, co-authored the paper with Berrueta and Murphey.
The study addresses the disconnect between traditional algorithms used in disembodied AI systems and robotics, highlighting the need for robots to collect high-quality data on their own. MaxDiff RL commands robots to move randomly to gather diverse data about their environments, enabling them to acquire the necessary skills to perform tasks effectively.
Through computer simulations, the researchers found that robots using MaxDiff RL learned faster and performed tasks more consistently and reliably compared to other models. These robots often succeeded in performing tasks correctly on the first attempt, even without prior knowledge.
MaxDiff RL is a general algorithm that can be applied to various applications, with the researchers hoping it will address foundational issues in the field and lead to reliable decision-making in smart robotics.
The study was supported by the U.S. Army Research Office and the U.S. Office of Naval Research.
