How robots learn from humans

The use of robots in domestic settings and in Industry 4.0 requires a significant degree of adaptability to manage the constant changes in the robot’s work environment. The classic paradigm used in robotics to execute complex tasks involves:

• Sensors to “perceive” the environment, paired with computational ability to develop an action plan;
• Actuators to carry out the planned task.

This paradigm has major limitations, requiring a unique system for each specific operational context, such as serving a meal to a person sitting at a table or lying in bed. Additionally, implementing this approach demands the integration of diverse software and hardware components along with extensive multidisciplinary expertise—all factors that limit the deployment of robots in highly dynamic environments like home settings.

“One way to increase the flexibility of robots,” explains researcher Matteo Saveriano, “is to enable them to learn new behaviors by using data collected from observing humans performing the same task. When properly trained, specific artificial intelligence algorithms can perform the task in various situations. While these algorithms are very effective, their training takes days and requires large volumes of data, which must often be processed manually. Furthermore, predicting the behavior of highly complex algorithms, like modern neural networks composed of millions of neurons, is challenging in a deterministic way.”

The Current State of Research

“At the Department of Industrial Engineering, we’re exploring the possibility of robots learning new tasks from humans, using algorithms that can train robots within seconds and with relatively little data. This allows us to quickly teach a single robot to perform multiple tasks. We’ve also developed various techniques to combine simple tasks into more complex ones. For instance, grasping a water bottle and opening it are both necessary steps to pour water into a glass,” Saveriano continues.

Our research aims to ensure that the artificial intelligence algorithm operates in a deterministic and predictable manner, viewing it as the evolution of a dynamic system and using a mathematically rigorous approach to guarantee closed-loop stability.

“These mathematical guarantees are essential for certifying the AI algorithm and using it to control a robot that operates closely with humans,” concludes Saveriano. “In the future, we hope to implement increasingly efficient AI algorithms to further enhance the ability to generalize a learned task, providing users with an increasingly natural interaction experience with the robot.”