Local
Tufts researchers test compact AI that helps robots navigate changing surroundings
A Tufts robot used a small, on-board AI system to steer around an obstacle it had not encountered before. Researchers say similarly compact models could cut energy use for some tasks without relying on remote data centers.
A wheeled robot named Ava navigated around a student’s desk in a Tufts University hallway, adjusting its route because the obstacle was not on its map. The demonstration illustrates how researchers are building smaller artificial-intelligence systems that can help robots respond to changing surroundings without connecting to a large data center.
Matthias Scheutz, director of Tufts’ Human-Robot Interaction Lab, said robots could eventually take on tasks ranging from search-and-rescue work to household chores. Ava, a white, camera-equipped robot, was asked to lead him to an elevator. When it met the unexpected obstacle, it stopped, then planned a path around it and continued.
The robot used a neurosymbolic AI model, which combines task execution with pattern recognition to help it respond to its environment. Because the model runs on a computer chip installed in the robot, it can make decisions without sending information to a remote server or requiring an internet connection.
That approach contrasts with many popular AI systems, which depend on large models and extensive computing infrastructure. Data centers require substantial electricity and water, including resources for cooling. Smaller models may reduce some of those demands, although their capabilities can be more limited than those of larger systems.
In a Tufts research test, the team put its neurosymbolic model on a robotic arm and compared it with a larger AI system on a block-stacking puzzle. The researchers reported that the smaller model was more accurate and used 100 times less energy to train. The team presented the results at an international robotics conference this year.
“Who doesn’t want a model that’s more capable given the same data or a model that’s as capable but requires less data?” said Helen Lu, a Tufts doctoral candidate and co-author of the study.
The work is part of a broader push to make AI more efficient. Liquid AI, an MIT spinoff, is developing compact models that can run on devices equipped with computer chips. The company has partnered with Mercedes-Benz in North America, and its chief science officer, Alexander Amini, said the goal is to find the “most intelligent model per watt.” He described potential uses such as a car responding to a driver’s request to ease back discomfort by activating a seat heater or massage function.
Smaller models may not match the breadth of large AI systems, but researchers say many everyday tasks do not require that level of computing. Daniela Rus, Liquid AI co-founder and director of MIT’s Computer Science and Artificial Intelligence Laboratory, said not every AI problem needs “cloud-level intelligence.”
The energy savings reported in the Tufts test apply to training the model; the researchers’ broader argument is that running AI directly on devices can also avoid the ongoing power demands of sending requests to and operating large data centers. The approach does not mean large AI systems will disappear, but it could offer a lower-resource option for particular robotic and consumer applications.