The new framework, TeNet (Text-to-Network), converts a command into a compact control neural network once, after which the robot can respond independently and at lightning speed. The research results will be presented during the international AI conference IJCAI-ECAI 2026.
Large Language Models (LLMs) make it possible to control robots using natural language. In practice, however, these systems are often too large and computationally intensive to be used directly on smaller robots or in applications where a response must be given within fractions of a second.
TeNet
To solve this problem, computer scientists Ariyan Bighashdel and Kevin Sebastian Luck developed TeNet within the NWO-funded Tenet project. Instead of continuously running a large language model during control, TeNet uses the language model only once to generate a compact, task-specific controller. Subsequently, the robot executes all actions using this lightweight control program.
The researchers demonstrate that this approach is both fast and efficient. The generated controllers contain approximately 40,000 parameters, significantly fewer than the 1 to 39 million parameters typical of existing models. As a result, TeNet achieves control speeds of more than 9 kHz, allowing robots to respond very quickly to changes in their environment and making them suitable for real-time applications.
Compact AI opens the door to a new generation of smart robots
The results could significantly broaden the deployment of language-controlled robots. Because the required computing power is much lower, smaller and cheaper robots can also utilize advanced AI without heavy or expensive hardware. This makes the technology more accessible for applications in, for example, healthcare, logistics, industry, and services.
Additionally, the new method offers advantages in terms of energy consumption. Because only a compact controller remains active during use, much less computing power is required than with systems that continuously run a large language model. This can lead to more energy-efficient robots and a more sustainable use of AI.
According to Luck, the efficiency does not come at the expense of language comprehension. “TeNet remains robust when users formulate the same command in different ways. As a result, people can communicate with robots in a natural way, without having to use fixed or pre-programmed commands.”