Stony Brook researchers have revealed a novel method that enables robots to keep learning over their lifetimes. This way, owners can teach their household assistants new tasks, trusting them to continuously refine their skills and performance.

Stony Brook, NY, Mar 3, 2025 — Imagine a robot assistant that not only helps around the house but gets smarter the longer it works with you. A group of researchers at Stony Brook and MIT has made a groundbreaking step toward turning this idea into reality. By developing a method for robots to continually learn and improve as they assist people over time, this research promises to make home robots more useful, adaptable, and intelligent. Their paper, titled ‘Embodied Lifelong Learning for Task and Motion Planning,’ appeared at the 7th Conference of Robotic Learning (CoRL).

Consider a home assistant robot spending its life in one house. It initially comes with a few basic capabilities — like moving an object from one place to another, rotating its hands, or pulling a door. Over its lifetime, the robot would be expected to acquire more experience, for instance, in learning how clothes are folded in this household, which books are kept where, and how to recognize the leftovers and put them in the compost box.

One of the challenges the researchers tackled was how the robot should handle different tasks and objects. For example, a robot might need to grab many types of objects, from a simple ball to a delicate vase. The study, led by Assistant Professor Jorge Mendez-Mendez, Department of Electrical and Computer Engineering at Stony Brook, presented an innovative solution.

Lifelong Robot Assistant learning to clean a living space

The idea was to design a system that helps the robot learn both general skills (like grabbing anything) and specialized skills (like grabbing a fragile object) while deciding, in the moment, which one to use. “Part of our solution,” added Jorge, “was to make sure the robot was learning new tasks more effectively. This is why instead of focusing solely on the final performance of a robot right after it was trained, we looked at how its capabilities were improving over time.”

Testing the system on sample 2D tasks and more realistic tasks from the BEHAVIOR  benchmark — a set of basic criteria for everyday household activities in virtual, interactive, and ecological environments — the team demonstrated that their approach leads to substantial improvements in performing new tasks more successfully. These results underscore the promise of lifelong learning systems for robots that continuously evolve based on their interactions with the world. “Our findings suggest that robots equipped with this kind of lifelong learning system will not only do their jobs better but will also become more personalized and intuitive over time,” Jorge added.

This research is a significant step toward creating robots that are not only capable of performing specific tasks but can also grow and adapt in response to the demands of their environments. “The method outlined in this study holds great potential for enhancing the intelligence of home robots, enabling them to offer more personalized, effective assistance over time,” Jorge added. The team’s work opens the door for future advancements, as robots learn and adapt to a wide array of tasks across their lifetimes, ultimately leading to a seamless integration of robots into daily human life.

Ankita Nagpal
Communications Assistant