Self-Debugging Robots: Fault recovery through reasoning and planning

Unexpected perturbations in an open-world task environment can cause various types of faults and failures which have to be dealt with to ensure long-term autonomous operation. We present an inference framework that enables an autonomous robot to generate and test fault hypotheses in open-world scenarios. The tests involve different types of introspective and overt behaviors based upon a derived fault hypothesis which informs behaviors which explore the failure. With suitable exploration, the robot can find the most sensible strategy to resolving the current failure condition if possible, allowing the task to be completed. We demonstrate the operation of our methods on a fully autonomous robot over a series of scenarios, ranging from a challenging yet solvable sensor occlusion case to mitigable or simultaneous failure cases.