Computer scientists at Vrije Universiteit Amsterdam have designed a new system that helps small robots understand voice and text commands without needing a heavy continuous connection to large language models.
The system, called TeNet, converts complex instructions into a compact neural network just once, allowing the machine to execute physical movements independently and in real time.
Cutting computing power
Large language models allow people to give robots instructions using natural spoken language, but these systems typically demand immense processing power and constant cloud connectivity.
That approach creates lag, making it difficult for automated machines to react instantly to sudden changes in their surroundings.
Researchers Ariyan Bighashdel and Kevin Sebastian Luck addressed this limitation by using the large language model only during the initial translation phase. Once the model generates a lightweight task controller, the robot runs entirely on that simple local script.
High speeds and low energy
The resulting controllers use roughly 40,000 parameters, a dramatic drop from the millions typically required by standard AI setups. This reduction enables processing speeds over 9 kHz, which allows robots to respond to real-time physical inputs within fractions of a second.
Because the compact controller runs locally, the technology requires significantly less power and opens up new possibilities for low-cost hardware. The breakthrough follows broader local industry shifts, such as how autonomous robots lay bricks on Amsterdam construction sites.
Lower hardware and energy demands could make voice-controlled automation practical for smaller businesses, healthcare facilities, and logistics centers across the Netherlands. The Amsterdam team will present their full findings at the upcoming IJCAI-ECAI 2026 artificial intelligence conference.
