It learns from the original

Innovative rat robot inspires researchers

The Munich neuro-robotic mouse is a kind of precursor to SMuRo.

(Bild: AP ( via APA) Austria Presse Agentur/APA/BING ET AL./SCIENCE ROBOTICS)

A Chinese-German research team has presented a robot modeled on a rat that learns the behavioral patterns of living "conspecifics" in the journal "Nature Machine Intelligence". The "SMuRo" system can interact with rats for half an hour, nudges them with its nose and shows other species-typical behavioral patterns.

For Graz researcher Thomas Schmickl, the work is remarkable for various reasons, as he writes in a commentary.

The idea of using robots that fit into a group as naturally as possible, for example, to conduct behavioral research, to guide groups of animals to a certain extent or to gain new insights into the interaction between animals, humans and artificial intelligence (AI) systems, is tempting for many researchers. However, as the team led by Qing Shi from the Beijing Institute of Technology (China) writes in their paper, it is technically very difficult to make such systems behave more or less authentically, to be accepted by their living role models and to elicit reactions from them. If this succeeds, we would be dealing with a "biocompatible" or "bioeffective" robot, as Thomas Schmickl, head of the "Artificial Life Lab" at the University of Graz, explains in a perspective article on the publication.

"SMuRo" system learns through observation
The Chinese scientists have now built a flexible robot that looks like a rat and moves around on a base with wheels. This allows SMuRo to imitate typical movements. However, these were not programmed into the system, but are learned by observing real rats. With the help of machine learning, SMuRo processes what it has seen and experienced and makes its own sense of it, so to speak. The scientists, who include Zhenshan Bing and Alois Knoll from the Technical University (TU) of Munich, are therefore using an approach that is modeled on social or imitation learning in humans and animals.

The Munich-based group has been working for years on the development of robots that can implement the movements of mice as authentically as possible. The team presented its neuro-robotic mouse (NeRmo) with colleagues from China last year in the journal "Science Robotics". The current SMuRo system is based in part on this work.

Robo-rat in "behavioral dialogue" with real animals
By observing rats interacting with each other, the robot gradually acquired knowledge about its "conspecifics". It collected data about the spatial position of the body parts during movements or about the way the animals behaved in a room. Based on this, the new system developed his inner picture of rats. This allows it to actively and authentically imitate movement sequences. What's more, the robot's actions demonstrably influenced the behavior of the rats in contact with it.

This is the first time that the social learning cycle has been demonstrated in a robot that interacts freely with animals, writes Schmickl. SMuRo demonstrates unprecedented social learning ability in a long-lasting "behavioral dialogue" with real rats, according to the Graz-based scientist, who recently wrote an article in "Science Robotics" about the possibilities of teamwork between robots and animals in the wild.

Expert: "Impressive step forward"
In the case of SMuRo, it was shown that the robotic animal learned the typical rat behaviors of holding, pouncing and social nose contact and used them to motivate its counterpart to engage in more exploratory behavior and encourage their curiosity. If the robot held the animals more tightly, they made more negative vocalizations, while increased jumping and nose contact resulted in positive feedback.

For Schmickl, "an autonomous robot that can cope with such long and complex interaction patterns" is an "impressive step forward" - especially when you consider how mercilessly animals reject a bad imitation of themselves: "It will be interesting to see where the door that SMuRo has opened will take us." Such "biohybrid systems", as the authors of the study put it, could help in future to understand "complex interactions between sensory perception, behavioral decisions and internal states" and to draw conclusions about the underlying brain functions. Ultimately, it could also be used to improve interactions between humans and AI systems, according to Qing Shi's team.