By Kimberly Mann Bruch
Breakthrough method developed by HDSI researchers helps artificial intelligence adapt to new challenges by breaking down problems into familiar building blocks
A team of University of California San Diego researchers have developed an artificial intelligence system that mimics how humans tackle unfamiliar problems — by combining pieces of knowledge they already have in new ways. The innovation could lead to robots and AI systems that are far better at handling unexpected situations.
“The challenge with robotic behavior is when AI meets an unfamiliar situation,” said Biwei Huang, an assistant professor with the Halicioğlu Data Science Institute (HDSI), within UC San Diego’s new School of Computing, Information and Data Sciences, which aims to advance data science and AI education. “Current AI systems, particularly those used in robotics, often struggle when they encounter environments or tasks they haven’t been specifically trained for and then it is like teaching someone to drive only on highways and expecting them to navigate a busy city center — the fundamental skills are there, but the new context throws everything off.”
Huang said that this limitation has been a major roadblock in developing truly versatile AI assistants and robots that could work reliably in unpredictable real-world settings. She said that she and Xinyue Wang, an HDSI graduate student, have been working on a novel system called World Modeling with Compositional Causal Components (WM3C) that takes a page from the human playbook.
“When people face new situations, they don’t start from scratch — they mentally break down the problem into familiar components and recombine them in novel ways,” Huang said. “Think about how you might approach cooking a dish you’ve never made before — you draw on your knowledge of individual cooking techniques, ingredient properties and flavor combinations to create something new.”
She explained that the WM3C system does something similar by learning fundamental building blocks, understanding connections and smart recombination. That is, the system identifies basic cause-and-effect relationships in its training environments and determines how these building blocks interact with each other. When facing a new situation, WM3C creatively combines these known elements rather than trying to learn everything from scratch. The system even incorporates language processing to help organize and label these building blocks, making them easier to identify and use appropriately.
“We tested our approach on both computer simulations and actual robotic tasks involving object manipulation,” Huang said. “The results were striking: WM3C significantly outperformed existing methods when robots were asked to handle objects or situations they had never encountered before.”
This success suggests the system could be valuable for developing robots that work in unpredictable environments — from household assistants that adapt to different homes to industrial robots that can handle varying manufacturing conditions.
“Our work continues, and we are making additional progress in creating AI systems that are more flexible and robust — potentially reducing the need for extensive retraining every time a robot encounters something new,” Huang said. “Instead of being brittle specialists, future AI could become adaptable generalists — much more like human intelligence.”
The research was presented earlier this year at the Thirteenth International Conference on Learning Representations.
