An international team of scientists combined several AI-based models and taught robots to learn new skills. In the future, devices will be able to serve themselves.
A team of researchers from Edinburgh and Zhejiang Universities has developed a way to combine deep neural networks to create systems with a new kind of learning. Scientists described the new architecture and its performance in the journal Scientific Robotics.
The researchers explained that, usually, deep neural networks can learn new movements and functions by learning repeatedly from the same examples. Such models are used in a wide variety of applications, face recognition systems or decision-making on a bank loan. The researchers combined several models designed for different skills to create an AI-powered super-system in the new work. This is how the model learned to learn new skills.
To do this, the researchers trained several deep neural models in various functions. For example, one of them learned to walk, and the second – to avoid obstacles. All the systems were then connected to a neural network, which over time learned to call other models that require special skills. The resulting system was able to fulfill all the skills of the combined models.
Model-based robots have learned dozens of skills on their own through trial and error. For example, the devices have learned how to get up properly after falling on slippery floors or what to do if one of the motors is broken. The researchers speculate that their work marks a new milestone in robotics research – now people won’t even interfere and fix device problems.