A Unitree G1 robot used vision language models to map indoor wireless signals

A new research paper details how a Unitree G1 robot used vision-language models to improve how it understands wireless signals in an indoor space. When a robot builds a digital map of its surroundings, it needs to know how different materials like wood, glass, or metal affect wireless signals. This process is called electromagnetic digital twinning.

Usually, robots struggle to know these material properties without manual setup or slow, random movement. The research team connected a vision-language model to a simulation platform called NVIDIA Sionna to automate this calibration.

The system performed two key tasks using the robot's camera. First, it identified building materials in the environment. It then mapped these materials to specific conductivity values to help the simulation software understand the room better.

Second, the robot planned its own path to take the most useful signal measurements. Instead of wandering randomly, it navigated to spots that would most effectively reduce errors in its digital model. The robot traveled less than 20 meters to achieve high accuracy.

Compared to random exploration, the system was twice as efficient at gathering the necessary data. It is quite a relief when the robot decides where to go instead of just bumping into walls until it gets lucky.

Have you ever wondered how your devices manage to stay connected while you walk between rooms with different types of walls?

This framework has specific limitations. The tests were confined to a controlled indoor setting and relied on the robot's ability to clearly see the materials in its environment. It has not yet been proven in outdoor areas or in spaces with highly reflective or complex moving objects.