Models pave the way for new applications in autonomous robotics
Researchers of UAntwerp and Flanders Make looked at nature to develop new sonar technologies
Researchers at the University of Antwerp and Flanders Make introduce two new groundbreaking technologies in sonar and echolocation research. The models showcase the potential of sonar as a powerful tool for robotic navigation and biological research.
The researchers at the Faculty of Applied Engineering (UAntwerp) and Flanders Make, the strategic research center for the manufacturing industry, started from the observation that current sonar sensors are limited in the amount of information they extract from the environment. Most of the current state of the art systems only extract the location simple objects from the acoustic data. In turn, this limits the amount of complex interactions that robots using this information can make with their environments, and limit the robots to do very simple tasks.
To overcome this issue, the researchers have looked at nature to better understand how to work with acoustic data. In biology, it is a common hypothesis that the brain is a prediction machine, constantly predicting the way the world should look like when perceived through the sensor organs (eyes, ears, touch, etc). This approach is called predictive processing. These predictions are made using a model of the world, which is refined based on the errors between the predictions and the observations.
Navigating in complex environments
Using the newly developed EchoPT model, which is like a “Chat-GPT” for sonar data, it becomes possible to predict how the sensor data should evolve over time while a robot is doing certain motions. Discrepancies between the predictions and observations can then be used to infer the state of the world, and make the robots much more robust when navigating in complex environments such as noisy industrial sites.
Accompanying to the EchoPT model, the researchers have developed a new model called SonoNERFs. Using this model, it becomes possible to reconstruct detailed 3D models of the world using the kind of sensor data available to echolocating bats. It uses the same principle of predictive processing as the EchoPT model. It serves as a base model of how bats could make complex decisions in complicated situations, and inspired the development of the EchoPT model.
Self-learning solutions
“These two new models showcase the potential of sonar as a powerful tool for robotic navigation and biological research, especially when combined with novel and powerful AI techniques” says Jan Steckel, a professor at the University of Antwerp. “EchoPT and SonoNERF offer self-learning solutions that go beyond what conventional sensors can achieve and bring us closer to understanding natural echolocation as seen in bats. With these breakthroughs, we hope to pave the way for new applications in autonomous robotics.”