With U.S. AI and data center energy use expected to double by 2030, researchers at the MIT School of Engineering have developed a proof-of-concept for AI systems that are up to 100 times more energy-efficient than current models while providing more accurate results on complex tasks.
Led by Matthias Scheutz, Karol Family Applied Technology Professor, the team employed neuro-symbolic AI, a hybrid approach combining conventional neural networks with symbolic reasoning similar to human problem-solving. The research will be presented at the International Conference of Robotics and Automation in Vienna this May.
Unlike large language models (LLMs) like ChatGPT or Gemini, which are screen-based, Scheutz’s team focuses on visual-language-action (VLA) models for robots. These models interpret camera and language inputs to generate real-world actions, such as manipulating robot arms or moving objects. Traditional VLAs are resource-intensive and prone to errors when interpreting complex scenes.
In testing with the Tower of Hanoi puzzle, the neuro-symbolic VLA achieved a 95% success rate, compared with 34% for conventional VLAs. For previously unseen, more complex puzzles, it scored 78%, while standard VLAs failed every attempt. Training time was reduced from over a day to just 34 minutes, using only 1% of the energy required by conventional VLA models. Execution of tasks required only 5% of the energy.
“Current AI systems are consuming extraordinary amounts of energy—sometimes more than entire small cities,” said Scheutz. “Our neuro-symbolic approach shows that AI can be both more reliable and far more sustainable, offering a path forward amid growing data center demands.”
The team’s findings highlight a potential solution to the growing environmental and economic concerns posed by large-scale AI, suggesting that hybrid AI systems may provide both efficiency and accuracy for industrial and research applications.
