AI and the science of abstraction

Artificial intelligence, or AI, promises transformative innovation for transportation, manufacturing, health care and education. It may also bring freedom from tedious tasks. Imagine robots doing laundry at your home or inspecting cargo at your local airport. These scenarios are not yet reality because of a longstanding problem in the field of computing: how to manage uncertainty.

“What should a household robot do if it finds a pet cat napping in a heap of dirty clothes? Or what should an inspection robot do with an unidentified package?” asked Siddharth Srivastava, an assistant professor of computer science in the Ira A. Fulton Schools of Engineering at Arizona State University. “Constantly asking humans for help is counterproductive, and immediate communication may not always be possible. We need them to compute what to do and to fall back to humans only when necessary.”

These examples represent what are known as open-world environments, and they are a world apart from the controlled conditions of a robotics lab. Robots and autonomous agent systems, such as Siri or Alexa, currently lack the ability to process unknowns, and relationships among those unknowns, to navigate open-world settings in a way that humans do intuitively.

Srivastava and his team at the Autonomous Agents and Intelligent Robots, or AAIR, laboratory research how sequences of decisions are made across extended periods of time amid uncertainty. In that domain, they are working to solve the problem of unknowns within AI and advance robotics to a new level of utility, reliability and safety.

“What excites me the most is determining how we can design algorithms that let AI systems automatically compute what they should do next, and then next after that, in order to reliably achieve complex, multi-step, user-assigned objectives in the real world,” he said.

Srivastava’s vision has captured the attention of the National Science Foundation, which has selected him for a 2020 Faculty Early Career Development Program (CAREER) Award. Such recognition is reserved for researchers who show the potential to be academic role models and to advance the missions of their organizations. CAREER awards provide approximately half a million dollars over five years to further each recipient’s research.

For Srivastava and his team, the key to success in this work is abstract — literally the ability to reason with abstractions.

“For instance, asking an autonomous system to bring me a cup of tea is an abstract instruction,” he said. “I don’t specify where the tea is located, how it should be made or where it should be brought. Lack of such detailed information can be viewed as an abstraction.”

He explains that the seemingly simple act of a robot delivering that hypothetical cup of tea involves thousands of decision points related to planning and movement through an uncontrolled environment. What if there are children running through the house? What if the power goes out? Humans manage these uncertainties without even thinking. But the actual process represents a distillation of critical information from a vast field of data.

“Finding concrete solutions in dynamic situations with unknown numbers and types of objects is difficult,” Srivastava said. “But in some of my earlier work, I found that identifying the right abstractions enables you to compute generalized plans or generalized solutions that work very efficiently. So, my group and I are using these methods to develop AI systems that can operate reliably and efficiently in open-world environments.”

Srivastava believes that the Fulton Schools community has the talent and resources necessary to develop the framework and the algorithms that will clear one of the biggest hurdles in AI. Through such innovation, doing the laundry may never be the same.