“Hey, Ava.”
Ava, a wheeled white pillar with a camera and screen docked in a Tufts University lab hallway, responded with a two-tone bleep.
Mattias Scheutz, director of the school’s Human-Robot Interaction Lab, continued his request.
“I’m wondering if you could take me to the elevators,” he said.
“Heading there,” the robot said from some hidden speaker. And the experimental pylon started to move.
Scheutz said he imagines a future where robots like Ava do many things big and small, from search-and-rescue missions to washing dishes. With recent advances in artificial intelligence, he said, that vision feels closer than ever.
This story is part of “Awaiting Data,” a special New England News Collaborative series about why there’s no AI data center boom in New England — yet.[/keyfigures]
Scheutz is one player in a cast of Boston-area researchers and engineers using AI to build robots that can adapt to changes in their physical surroundings. And the tech solution they’re harnessing doesn’t rely on data centers.
Popular chatbots like Claude and ChatGPT are powered by models trained on enormous amounts of information, which means spending billions on vast data centers that can take more power and water than entire cities. Instead, roboticists like Scheutz are thinking small. They’re designing AI systems compact enough to fit on a single computer chip.
Robots powered with these small AI models can make decisions quickly. And because the chips are installed in the robots, there’s no need to send signals to and from data centers for processing. That means the robots can run AI without an internet connection.
On its journey to the elevators, Ava encountered a student sitting at a desk in its path. The robot stoped and stuttered: It’d never seen this obstacle before.
“That’s not in its map,” Scheutz explained. “So now it has to plan around it.”
Within seconds, Ava skirted around the desk and continued onward.
By making that little adjustment, Scheutz said Ava accomplished a key challenge in robotics; it navigated what’s known as a “dynamic obstacle.”
The robot tackled this task with a type of AI called a neurosymbolic model. The model allows the robot to execute tasks and recognize patterns in its environment, which allows Ava to move in response to unfamiliar situations.
Models like this reduce the most energy-intensive aspect of AI: training. To create the large language model GPT-4, for example, MIT Technology Review cited an estimate that Open AI used as much energy as San Francisco consumes in three days.
And small AI models have another energy-saving bonus. Since they don’t need data centers to operate, they bypass the energy costs of running hefty cooling systems and thousands of servers at once.
The Tufts team loaded its neurosymbolic model onto a robotic arm and tested it against a larger AI model. The researchers asked both systems to tackle a block stacking puzzle. The results, which they presented at an international robotics conference this year, showed the neurosymbolic model was more accurate and required 100 times less energy to train, compared to the larger AI.
“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?” asked Helen Lu, a Ph.D candidate and co-author of the Tufts study.
Other companies are working on their own versions of small, energy-efficient AI.
“We want to find the most intelligent model per watt,” said Alexander Amini, chief science officer at Liquid AI, an MIT spinoff company.
Amini said the startup’s strategy runs counter to most big AI companies, which focus on creating intelligence “at all costs.” By honing in on smallness instead, Liquid AI can run its models on any device with a computer chip.
Like the chips in your car. The company has partnered with Mercedes-Benz in North America, and Amini envisions cars that respond to conversational prompts like, “My back hurts.”
“And the car is intelligent enough to reason about those statements and start to, let’s say, turn on the seat heater, turn on the massager,” he said.
A potential downside is that limited training may create AI models with less intelligence, and less usefulness, than large AI. But Liquid AI’s cofounder, Daniela Rus, argues there’s plenty of room for both.
“It is important to keep in mind that not all AI problems require cloud-level intelligence,” said Rus, who directs MIT’s Computer Science and AI Laboratory.
A lot of what users ask AI chatbots are relatively simple queries that can be answered without “extreme Ph.D-level knowledge,” according to Amini.
Big AI — the type of model that can both do astrophysics and write a book report — isn’t going away. But Rus and Amini are betting that not every task, or every robot, needs that.
____
This story is part of Awaiting Data, a special series from the New England News Collaborative.
This article was originally published on WBUR.org.
Copyright 2026 WBUR
