For future robotics, knowing how a robot will respond under different conditions is necessary to strengthen safe operations. The complexity is with assessing understanding what might break a robot without actually carrying out the activity and damaging the machine. The answer lies in a a new machine learning method. The method, which comes from the Institute of Science and Technology Austria, can make use of observations gathered under safe conditions.
The algorithm can then make accurate predictions across a range of possible conditions, based on the same physical dynamics, allowing a robot to predict which activities would be safe and which would be dangerous. This provides a resemblance of the cognition that human shave in sizing up risks before an action is taken. This requires extrapolation, or making predictions about situations outside of the known.
The approach used by the researchers is based on a shallow neural network approach, which enables an understanding of functional relations. This creates a class of learnable equations for a learning network to comprehend. The criteria adopted by the researchers was that the model must be plausible, which they defined as consisting “of components that have physical expressions in the real world”. Further, the model should be interpretable: “which typically means that they consist only of a small number of interacting units.”
This isn’t just theoretical, according to lead researcher Georg Martius: “We’re actually working on developing a robot that uses this type of learning. In the future, the robot would experiment with different motions, then be able to use machine learning to uncover the equations that govern its body and movement, allowing it to avoid dangerous actions or situations.”
The research has been presented as a white paper to the 35th International Conference on Machine Learning. The paper is titled “Learning Equations for Extrapolation and Control.”