Large language models have made it possible to control robots using natural language, but in practice these systems are often too computationally intensive for smaller robots or applications that require responses within fractions of a second. TeNet addresses this by using the language model only once: it converts an instruction into a compact neural network for robot control, after which the robot operates independently at high speed.

The framework was developed by computer scientists Ariyan Bighashdel and Kevin Sebastian Luck within the NWO-funded Tenet project. The generated controllers contain approximately 40,000 parameters, significantly fewer than the 1 to 39 million parameters common in existing models. This compactness allows TeNet to achieve control speeds of more than 9 kHz, enabling robots to respond rapidly to changes in their environment and making them suitable for real-time applications.

The reduced computational requirements also have implications for the accessibility of the technology. Because no heavy or costly hardware is needed, smaller and less expensive robots can also benefit from advanced AI. This makes the method potentially applicable in sectors such as healthcare, logistics, industry, and services. In addition, a system that keeps only a compact controller active during operation consumes considerably less energy than systems that continuously run a large language model, contributing to more sustainable use of AI.

According to Luck, the efficiency does not come at the expense of language comprehension. TeNet remains robust when users phrase the same instruction in different ways, allowing people to communicate with robots naturally without relying on fixed or pre-programmed commands.