Researchers at TU Wien have developed a method that allows robots to learn tasks such as cleaning or sanding through imitation, similar to humans. This involves collecting data from a sponge equipped with force sensors and tracking markers, which is then fed into a neural network.
Robots make our daily lives easier. In industry, they are used to reliably perform repetitive tasks. However, it becomes challenging when the shape of the workpiece changes or the task needs to be performed with varying intensity. If a robotic arm is supposed to clean or paint an object with corners and edges, it is difficult to precisely fit these tasks into fixed rules and predefined mathematical formulas, as described by researchers at TU Wien. For instance, a cleaning robot must consider that the sponge needs to be used at different angles and with varying force.
For such applications, the researchers use an approach where robots learn through imitation and experience, similar to humans. A person cleans a sink with a specially equipped sponge. The force sensors and tracking markers on the sponge generate data, which is statistically processed and used to train a neural network. “The robot learns that the sponge must be held differently depending on the surface shape,” explains doctoral student Christoph Unger, “and that more force is needed on a highly curved area than on a flat surface.” The key is that the robot learns this through a novel learning algorithm, even though only part of the sink was cleaned during the demonstration.
Looking to the future, the researchers see potential in mobile robots that work in smaller areas like workshops and exchange their collected experiences with each other. In this “federated learning” process, the basic knowledge of the respective applications would be shared with other robots. Elsewhere, researchers are already working on helpful robot prototypes that interact with humans through speech.
This new method of teaching robots through imitation and experience opens up new possibilities for their use in various fields. By observing human actions and translating them into robot actions, robots can become more adaptable and efficient in performing tasks. This development is especially promising for tasks that require a high degree of precision and adaptability, such as cleaning complex surfaces or performing delicate operations.
The use of neural networks in this context allows robots to learn from a limited set of examples and generalize this knowledge to similar tasks. This means that once a robot has learned how to clean a specific type of sink, it can apply this knowledge to other sinks with similar features. This reduces the need for extensive programming and allows for more flexible robot deployment.
Furthermore, the concept of federated learning among robots could lead to rapid advancements in robot capabilities. By sharing knowledge and experiences, robots can collectively improve their performance and adapt to new challenges more quickly. This collaborative approach could significantly enhance the efficiency and effectiveness of robotic systems in various industries.
Overall, the research conducted at TU Wien represents a significant step forward in the field of robotics. By enabling robots to learn through imitation and experience, we can create more versatile and intelligent machines that can assist humans in a wide range of tasks. This not only has the potential to improve productivity and efficiency but also to open up new opportunities for automation in areas that were previously too complex for robots to handle.