Tufts researchers test lower-energy reasoning model for robots

AI systems and data centers used about 415 terawatt hours of power in 2024, accounting for more than 10% of U.S. electricity production, according to the International Energy Agency. With demand projected to double by 2030, researchers at Tufts University are testing a proof-of-concept system meant to reduce the energy burden of AI while improving reliability.

The system uses neuro-symbolic AI, combining neural networks with symbolic reasoning to help visual-language-action models in robotics break tasks into rules, categories and steps. The approach is designed to avoid the heavy trial-and-error learning used by conventional systems, which can misread objects, mishandle physical tasks or produce errors similar to hallucinations in large language models.

In tests using the Tower of Hanoi puzzle, the neuro-symbolic VLA achieved a 95% success rate, compared with just 34% for standard systems. On a harder version it had not seen before, the hybrid system succeeded 78% of the time, while traditional models failed every attempt. It learned the task in only 34 minutes, required only 1% of the training energy used by a standard VLA system, and used just 5% of the energy during operation.