Active Learning Omnivariate Decision Trees for Fault Diagnosis in Robotic Systems

Robotic systems are highly complex, and debugging failures in them can prove challenging. We propose a technique for using multivariate decision trees to create human interpretable descriptions of the input conditions that cause these failures in robotics systems. This approach uses active learning to efficiently create tests, and uses a multivariate decision tree that captures common boundary conditions in the software fault space. We provide an evaluation of this technique on a small set of faults from several robotics systems, and compare it against a previous technique in this space, Hierarchical Product Set Learning. Our proposed technique requires fewer tests, provides more accurate estimates of the fault conditions, and is more interpretable than the prior approach.